# Changelog

Dated record of design and firmware changes, newest first. The [README](README.md) describes the
current design only; this file is how it got there. Entries marked "not yet tested on hardware"
compile and pass the PC-side checks (ERC/DRC, and `tests/` for the speed control) but have not run
on a built board.

> The one-shot scripts named in the dated entries below (`cap_fixes*.py`,
> `add_c9.py`, `labels_and_c9_pcb.py`, `fix_q5_q6_*.py`, `widen_power.py`,
> `reroute_24v.py`, `swap_module_sockets.py`, `cleanup_after_route.py`, and
> the original `build_pcb_100.py` / `route_pcb_100.py` / `place_check.py`
> placement flow) were removed after being applied. They can't be re-run
> against the current board. They're in git history, last present in commit
> `b2d7025`.

- **2026-11-18 The GDTF fixture has a Grid mode (tools, docs).** A third DMX mode, **Grid**, for the Art-Net tab's Grid layout: each of the 178 pixels at its cell of the 31 x 15 grid, as one run of 1395 channels over three universes (breaks 1-3; LED 62's channels straddle the first two), the cells without an LED left out so a console patches only the ring. `gen_gdtf.py` reads the grid from `artnet.h` and refuses an LED off a cell or two on one. Checked: valid against gdtf.xsd as before (keyrefs separately), and every Grid channel is where the firmware's `gridCellOf()` puts it (a program printing the firmware's cell for each LED, compared channel by channel). Same fixture type ID, so a console updates it; a second Revision says so. The user guide, the documentation index and the Art-Net tab say it covers every layout.
- **2026-11-18 Fixture links go through the documentation index (firmware web pages, docs).** A browser blocks a download started from the unit's plain-HTTP pages, so the Art-Net tab and the unit's copy of the user guide now link the documentation index's *Fixture files* section (HTTPS, with the files' download links and what each is for) instead of the files, and so does the user guide everywhere (its Fixture files section names the files and links the index, which the unit's copy reaches on the documentation site). `gen_gdtf.py --check` no longer looks for the file name in the firmware.
- **2026-11-18 The Art-Net tab links the fixture files; the user guide has an Art-Net chapter (firmware web pages, docs).** The Art-Net tab now shows links (each opening a new window) to the GDTF fixture and the two Resolume files on the documentation site, and to the user guide's new chapter; `gen_gdtf.py --check` fails if the link doesn't name the current GDTF file. The new-window icon moved to `web_theme.h` with the site's address (`DOCS_SITE`). The user guide's Art-Net text, until now a run of paragraphs inside Lighting, is chapter **8. Art-Net and sACN**: Turning it on (four steps), Layouts (with the pairs diagram), What the ring shows, Lighting desks find it (ArtPoll), Fixture files and Checking what arrives; Lighting points to it, and the chapters after it move up one (Network setup is 9 ... Specifications 13), every link to them updated.
- **2026-11-18 Art-Net settings get their own tab; the fixture files are linked (firmware web pages, docs, demo).** Receiving, the first universe and the layout move from the LEDs tab to a new **Art-Net** tab of the settings page (`#artnet`, in the menu's Settings list after LEDs; `ledAppendArtnetPanel()`); they are saved with the rest, the settings keys unchanged. The Status tab's Art-Net / sACN box points there. The fixture files are linked from the documentation index (a new *Fixture files* section) and the user guide, and the unit's own copy of the user guide links them on the documentation site (opening in a new window), as the unit doesn't serve them; the Art-Net tab's help does too. The user guide gains an Art-Net tab section.
- **2026-11-18 Art-Net grid layout, for one Resolume fixture (firmware, firmware web pages, tools, tests, demo, docs).** Resolume's fixtures are rectangles, every cell sending channels, so the ring took 20 slices of 11 strip fixtures. The LEDs tab's **One universe** box is now a **Layout** choice: **Two universes** (as before, the default), **One universe** (as before) and new **Grid**: the ring as one rectangle of 31 x 15 pixels, left to right, top to bottom, cells 6.35 mm across and 12.7 mm down so every LED (on its brick pattern) is exactly on a cell of its own, the rest ignored: 1395 channels as one run over three universes, a cell straddling two where it falls (`artnet::gridCellOf()`, `gridSlotOf()`, `applyGrid()`; setting `artgrid`, in backups). The receive path, sACN joins, ArtPollReply (three ports) and the Status tab follow the layout; `/netmon` gains `layout`. New `tools/gen_resolume.py` writes `docs/fixtures/resolume/`: the fixture (31 x 15 RGB) and an Advanced Output setup with one Lumiverse slice, sized so every cell's centre is an LED's place (1116 x 1080 px from x 402 in 1920 x 1080), from a template of the bench's Arena 7.23 setup; both match the bench's working files element for element. Tests: every LED on its own cell within 0.05 mm, a three-universe frame (a straddling cell included) lands right, other universes ignored. Not yet tried in Resolume: Auto Span is taken to send the channels as one run.
- **2026-11-18 A GDTF fixture for lighting consoles (tools, docs, CI).** New `tools/gen_gdtf.py` writes `docs/fixtures/Uncle Magic@Liquid Turntable@2026-11-18.gdtf` (and its `description.xml` unzipped, for diffs) from `src/led_map.h` and `src/artnet.h`: one fixture type (DataVersion 1.2, a fixed FixtureTypeID so a console updates it on re-import) with two modes, **2 universes** (178 RGB pixels, break 1 channels 1-510, break 2 channels 1-24) and **1 universe** (170 RGB elements, the eight shared pairs as Pair geometries holding their two pixels), and each pixel a Glow beam at its place on the ring (metres, as the pages draw it, facing up). Checked against the official gdtf.xsd (mvrdevelopment/tools, 2025-08-21): valid, except that .NET can't resolve the schema's keyref rules (keys on nested elements), so those were checked separately: every geometry, model, attribute and feature a channel or geometry names exists, every InitialFunction follows GDTF's naming, no channel is used twice or skipped, and the one-universe elements are the firmware's own (`singleElementOf()`). CI checks the file is current (`--check`). The user guide says how to use it.
- **2026-11-18 FAQ: two units, one scrambled (docs).** New entry under Lighting: two turntables on the same Art-Net broadcast where one shows the picture scrambled, as two did on the bench: their **One universe** settings differed. Why a mismatch scrambles (each LED after a shared pair takes a neighbour's colour), and to set both units, and the first universe, to match the software.
- **2026-11-18 The Others list restyled (firmware web pages, demo).** **Others** is a small outlined pill (violet while open or under the pointer), so it reads as a control beside the OSC address rather than another menu. The list has a header with the count; each turntable is a row with a dot (green for this one), its name, a tag (**this one**, **older firmware**, or **newer build** when its dated build is newer than this unit's) and a line for its address, OSC port and build; rows highlight under the pointer, and this one's is tinted. A footer says when the list was found (or that it is still looking) with **Search again**.
- **2026-11-18 The flips belong to the Art-Net effect; Others opens the turntables list (firmware, firmware web pages, tests, demo, docs).** The flips of the entry below are now part of the look, changed live, not LEDs-tab settings: keys `hflip=1` and `vflip=1` (written only when set), a **Flip** row (left-right, top-bottom) in the live page's effect settings and the LEDitor, shown for the Art-Net / sACN effect, so they are saved with a preset and take OSC `/look`. In `ledfx::render()` the artnet effect mirrors each LED's source place after turning it back (flipped, then turned): on this symmetric ring a mirrored place is exactly an LED and the next is 12.7 mm off, beyond the 12 mm reach, so a flip alone is exact without a lookup table (`makeMirror()` is gone), and the demo's engine does the same. The LEDs tab's checkboxes and `artfh` / `artfv` are gone; a backup that has them still restores (retired keys). Tests: a flipped still picture is each LED's mirror's colour, both flips are half a turn, a flipped picture turned a quarter is the mirrored picture turned a quarter, the keys read and write back, and two flipped looks in the demo parity check (24 looks). The menu bar's OSC address is a plain label again; the list of turntables opens from an **Others ▾** button beside it.
- **2026-11-18 Art-Net picture flips (firmware, firmware web pages, tests, demo, docs).** The LEDs tab has **Flip horizontally** and **Flip vertically** under the Art-Net settings (`artfh`, `artfv`; in backups' `[led]`): the received picture mirrored left-right (x to -x) and top-bottom (y to -y) as the pages' pictures show the ring, both together being half a turn. The ring's LED layout is symmetric both ways (every mirrored LED lands exactly on another), so `ledfx::makeMirror()` makes a reordering once at start and the render task shows LED i the colour received for its mirror, before the artnet effect turns and punches it: exact, nothing blended. Tests on the real LED map: each flip a reordering of all 178, twice is none, left-right moves x to -x, both is half a turn at the same radius, and the effect shows a flipped picture as it is.
- **2026-11-18 The menu bar lists the other turntables; desks see the unit (firmware, firmware web pages, tests, demo, docs).** The menu bar's *OSC address:port* now opens a list of the turntables on the network: this one first, then the others by name, each with its address on the browser's subnet, OSC port and build; picking one opens the same page on it. Each unit's mDNS announcement (`_http._tcp`, instance *Uncle Magic's Liquid Turntable*) now carries TXT keys `osc` and `build`; a unit with older firmware is still found, without them. New `GET /peers` answers at once from the last search and starts one (`mdns_query_ptr`, 2 s, its own task, one at a time) when that is over 20 s old, or with `?fresh=1` (**Search again**); a unit on both links is one entry with two addresses (`src/peers.h`, with `tests/test_peers.cpp`: which answers count, merging, order, JSON escaping, a full list). While Art-Net receiving is on, the unit answers ArtPoll with ArtPollReply (Art-Net 4, 239 bytes, `artnet::pollReply()`, tested field by field): one per port (the first universe, and the next unless one universe is ticked; BindIndex 1 and 2), unicast to the desk's port 6454 from the link on its subnet, with the network name as the short name, so desks list the unit; receiving off, polls are only counted. `/netmon` gains `artpoll` (polls, the last poller, replies sent), shown in the Status tab's Art-Net / sACN box. The demo lists two made-up neighbours.
- **2026-11-18 Preset 1 is built in: the user colour, always (firmware, firmware web pages, tests, demo, docs).** A unit's preset 1 showed a rainbow: it had been saved over (the live page's effect settings, Save to 1), and nothing stopped that. Preset 1 is now built in, as the factory palettes are (`presetOne()` in `src/ledpreset.h`: solid, the user colour, still, "User colour"), not kept in the file: `parse()` puts it in every table and skips any line numbered 1 (an older file or backup has one; it is not an error), `formatAll()` writes 2 and up, the built-in list (`kDefaults`) no longer has it, and a single save to 1 is refused, saying why (`ledPresetSaveLine()` now returns the reason, shown by the LEDitor and the live page). The live page's **Save to** list starts at 2 (and offers 2 while preset 1 runs); the LEDitor can still load preset 1 to copy it. A unit whose preset 1 was saved over shows the user colour again once updated, with nothing else to do; its file keeps the old line, skipped, until the list is next saved. Tests: the defaults read 22 presets plus the built-in one, a file whose preset 1 is a rainbow still gives the user colour, an empty file still has preset 1, and `formatAll()` starts at 2. The demo does the same.
- **2026-11-18 The Art-Net / sACN monitor moves to the Status tab (firmware web pages, demo, docs).** The live page's **Network** box is gone; all it showed (receiving on or off, the first universe and its Net, SubNet and Universe, whether the ring shows it, packets used, each universe arriving with its sender, packet count, age and whether this unit uses it, and other packets on the ports) is now the Status tab's **Art-Net / sACN** box, under **Now**, from the same `/netmon`, once a second while that tab is open and the page in view. The LEDs tab's help points there. The live page is four boxes.
- **2026-11-18 The live page and the LEDitor use wide windows (firmware web pages, demo, docs).** Their boxes now go into as many 480 px columns as the window has room for, one to four (`boxes()` in `src/web_box.h`; one column is the page as before, and the page widens to fit the columns), and dragging by the grip moves a box to any place in any column (the column under the pointer, the place by its height, as before), with a mouse or a finger. Each column count keeps its own arrangement, so a computer's columns don't change a phone's order: one column under the old key (`liveOrder`, `ledOrder`), so an order kept before still holds, and N columns under the key with N after it, as `box,box|box|box`. A column count with nothing kept starts from the one-column order, each box going into the shortest column so far. Tried in the demo at 1600, 1100 and 375 px: three, two and one column; a box dragged into another column, a finger drag at phone width, and the three-column arrangement back after a reload.
- **2026-11-18 One white background setting; /update's Upload in its box (firmware web pages, demo, docs).** The live page's **Background: white instead of off** box is gone: it did what each preset's own **Background** (`bg=white`, in the effect settings under it and the LEDitor) does, for every preset at once, until a restart. Gone with it: `/control` `white` (now refused as an unknown action), `/status` `white` and `lightWhite`. On `/update` the **Upload** button is inside the **New firmware** box, under the file picker, and every page's file picker (also the Backup page's) has its **Choose file** button drawn like the unit's other buttons.
- **2026-11-17 The unit says when a newer build has been released (firmware web pages, tools, docs, demo).** The Status tab's **Firmware** row and `/update`'s **Current:** line now say *Update available: <name>* (with a **Download** link to `latest.html`), *(up to date)*, or, offline, that they couldn't check. The browser showing the page asks the firmware site, not the unit, so the unit needs no internet or HTTPS: `tools/publish_firmware.py` now also writes `latest.json` (the newest build's name, as `FW_VERSION` reads, its file and size) and an `.htaccess` that lets other origins read that one file (`Access-Control-Allow-Origin: *`, `no-cache`). Build names sort by date, so the greater is newer; a build not made by `tools/build.ps1` just names the newest released. The check is `FW_UPDATE_CHECK()` in `src/version.h`, with `FIRMWARE_URL` (moved there from `ota.cpp`); the demo leaves it out. Tested from another origin against the live `latest.json`: an older build reads Update available, the same build up to date.
- **2026-11-17 /update marks its download links with a new-window icon (firmware web pages).** **Latest firmware** and **Archive** end in the usual box-and-arrow icon (inline SVG in the link's colour, labelled "Opens new window" for screen readers and on hover) instead of the text "(Opens new window)".
- **2026-11-17 /update reads "Current: <name>" on one line (firmware web pages).** It was **Current firmware:** with the build's name on the line below.
- **2026-11-17 The firmware row drops the partition name; /update shows the name alone (firmware web pages, demo).** The Status tab's **Firmware** row, `/status` `build` (so the About page) and a backup file's header now read `liquidTurntable_YYYY_MM_DD_HHMM (commit ...)`, without the OTA slot (`app0`/`app1`); the boot log still names the slot it runs from. `/update` shows **Current firmware:** with the build's name on the line below, without the commit.
- **2026-11-17 The firmware compiles in WSL (tools, README).** A build that added an `#include` took about 11 minutes: arduino-cli redoes its library scan from the first file whose `#include` lines changed, one preprocessor run per file, and on Windows each run took 8.8 s, 7.5 s of it in the kernel, as every `#include` probes about 400 ESP32 include folders (about 90 us a failed lookup on Windows, 3.5 us in WSL; Defender folder exclusions changed nothing). `tools/build.ps1` now compiles in WSL when arduino-cli is set up there (new `tools/setup_wsl_build.sh`: arduino-cli 1.5.2-rc.1, checked against its checksums, the ESP32 core 3.3.11, ArduinoOSC 0.6.0 and TMCStepper 0.7.3, as CI), and on Windows otherwise (`-Windows` forces it); flashing stays on Windows. Measured: a full rescan 61 s in WSL, an ordinary build 26 s (60 s on Windows), a build from nothing 100 s. The binaries differ only in the source paths in log and assert messages.
- **2026-11-17 The unit names its build as the archive does; /update's links say they open a new window (firmware web pages, tools, demo).** Under its heading the page says **Current firmware**: the build's name, as its distributable and the archive name it (`liquidTurntable_YYYY_MM_DD_HHMM`), and its commit. `tools/build.ps1` writes them to `src/fw_version.h` (git-ignored) before compiling (`src/version.h` reads them), and names the distributable from the same time; a build made otherwise (CI, the IDE) shows when it was compiled. The same name and commit replace the compile date on the Status tab's **Firmware** row, in `/status` `build` (so the About page), in the boot log's first line and in a backup file's header. **Latest firmware** and **Archive** each carry "(Opens new window)". The old `liquidTurntable_latest.bin` is deleted from the server, which now holds only `latest.html` and `archive/`.
- **2026-11-17 Firmware downloads from /update no longer blocked; no more latest copy (firmware web pages, tools, docs).** Chrome blocked the latest build's download as insecure: the file is served over HTTPS, but the click came from the unit's own page, which is plain HTTP. The link (now **Latest firmware**) opens `latest.html` on the firmware site in a new tab, which starts the newest archived build's download itself, with a button if it doesn't, so the download comes from an HTTPS page. `tools/publish_firmware.py` writes it with every publish (and `--index`), naming the build. With that page pointing at the newest file, `liquidTurntable_latest.bin` is no longer uploaded: a publish puts the build in `archive/` only. The Archive link opens in a new tab too.
- **2026-11-17 Release candidates are published, and /update links to them (tools, firmware web pages, docs).** `tools/build.ps1 -Publish` now uploads the build (new `tools/publish_firmware.py`, over SFTP with the docs upload's login and pinned host keys) to kennedy-williams.net/unclemagic/firmware/liquidturntable/: as `archive/liquidTurntable_YYYY_MM_DD_HHMM_firmware.bin`, kept, and as `liquidTurntable_latest.bin`, replaced each time (each uploaded under a temporary name and swapped in, so a download never gets half a file), with `archive/index.html` listing the archive newest first (the server doesn't list folders). Publishing is only ever asked for (a plain build doesn't), for a release candidate: `-Publish` refuses, before compiling, a tree with uncommitted changes or a commit not yet on GitHub, so every published build can be checked out; a failed upload is reported and the flash carries on. The distributable's name gains an underscore: `liquidTurntable_YYYY_MM_DD_HHMM_firmware.bin`. The `/update` page has a **Download firmware** box under the upload form linking to the latest build and the archive (downloaded by the browser; the unit fetches nothing). Also fixed in the READMEs: the build command read `toolsuild.ps1` (a backspace character where `` was).
- **2026-11-17 The GitHub repository is LiquidTurntable (README, docs).** The repository is renamed from `ESPMagicTurntable` to `LiquidTurntable` (github.com/scoobiekw/LiquidTurntable), to match the sketch and the project's name; GitHub redirects the old address. The READMEs name it, and `docs/sync_files.py` writes the new address into the published copies' links.
- **2026-11-17 The sketch is LiquidTurntable.ino, out of OneDrive (firmware, tools, CI, README).** `sketch_esp32s3_eth_v2.ino` is renamed `LiquidTurntable.ino` and the project lives in `C:\Software\LiquidTurntable` (Arduino wants the folder named as the sketch), out of OneDrive, which held new build files open as it uploaded them. The build's files are now `LiquidTurntable.ino.bin`, `.merged.bin` and so on; the distributable keeps its name (`liquidTurntableYYYY_MM_DD_HHMM_firmware.bin`). CI builds the sketch in a folder of that name.
- **2026-11-17 The menu's address says it is OSC's (firmware web pages, demo).** The address at the top left of every page now reads `OSC <address>:<port>` (where show software sends OSC), and before the unit has an address it shows `<host>.local:<port>`, not the name alone.
- **2026-11-17 One build script, with the old builds kept (tools, README).** `tools/build.ps1` is now the usual build: it compiles into a staging folder and, only if that succeeds, moves the previous build's firmware (`.bin`, `.elf`, `.map`) to `build/archive/<date>_<commit>/` (the newest 10 kept) and puts the new one in `build/`, with `build_info.txt` naming the commit (`+changes` if the tree wasn't clean) and the distributable, the app image as `liquidTurntableYYYY_MM_DD_HHMM_firmware.bin` (named with the build's date and time). `-Upload` flashes those same files over USB. OneDrive holds new big files open while it uploads them, so each move and delete is retried for up to 10 s (a move then becomes a copy). arduino-cli 1.5.2-rc.1 also copies each build's binaries to `build/esp32.esp32.esp32s3/` unasked; the script removes that copy. The README's network update steps point at it.
- **2026-11-17 The Art-Net effect starts where the sender put it (firmware, tests, demo).** Loaded, the artnet effect was turned by whatever angle the look before it had left (every look shares one rotation, so the others carry on smoothly from each other), and with speed 0 it stayed that far off, with nothing to bring it back. Now (`ledfx::alignNet()`, the demo's too) its turn starts at 0 when it is loaded (a new preset, or a change to it from another effect), and a still artnet look (speed and drift 0) is always shown unturned. Other effects carry on as before.
- **2026-11-17 A fast swipe back to where it started now lands there (firmware web pages, demo).** The hue slider swiped fast from 0 and back stayed part way (on any preset). The unit's request log showed every value it was sent arrive and apply in order, but the last was 0.80: the page sends a moving slider at most every 120 ms and left the stopping place to the slider's change event, which browsers don't fire when a drag ends where it began. Every slider on the live page now also sends where it stopped 150 ms after it stops, and the settings page's lowest platter speed does the same (and keeps it), 400 ms after.
- **2026-11-17 The Art-Net effect's turning made cheaper again (firmware, tests).** Turned, each LED now looks for its colour only among the LEDs in the 3 x 3 grid cells round the place it comes from (`ledfx::NetGrid`, cells 12 mm square, the reach: 16 x 15 on this ring, 7 LEDs to look at on average where the radius band of the entry below had 56), made once by `makeNetGrid()` and passed to `render()` (without one it looks at every LED, as before); the colours are the same (the test compares the two at 96 angles). The radius band is gone from `Geo`. Measured on the unit with sACN arriving and the effect turning a turn a beat: the slowest frame in each 2 s is now 3.9-6.1 ms (from 7.5-12.4 ms, and 10-13.6 ms before the entry below), a plain preset's 2.6-3.4 ms; 100 frames a second throughout. Tried and dropped: holding a faded-out Art-Net look's last frame for the half-second crossfade instead of drawing it every frame. On the unit it saved about 1.6 ms a frame for the fade (slowest frame 3.3-4.4 ms held, 4.2-5.9 ms drawn, both well inside the 10 ms), but the old picture visibly froze: run through the engine on a PC, a quarter turn a beat left 26-94% of the LEDs more than 8 levels off (up to 224 of 255) in the fade, and a sender animating its picture froze too.
- **2026-11-17 No more watchdog resets with the Art-Net effect turning (firmware, tests).** The unit reset (task watchdog: `ledTask` on core 0 for 5 s, the idle task starved) with the Art-Net effect running. The render task, a priority above the idle task, waits for its next 10 ms frame with `xTaskDelayUntil()`, which doesn't wait at all once a frame has taken its whole slot: with frames that long it never blocked. A late frame now still sleeps a tick and the frames count on from there, so whatever an effect costs the idle task runs. The cost: turned, the Art-Net effect looked for each LED's colour among all 178 LEDs (31,700 distances a frame, twice during a crossfade). Turning keeps the radius, so it now looks only among the LEDs within 12 mm of the LED's own radius (`Geo::byRadius`, `nearLo`-`nearHi`, made once in `makeGeo()`: 56 on average), skips those out of reach on x alone, and multiplies instead of dividing; the colours are the same (a new test against the full search at 96 angles). Measured on the unit with sACN arriving and the effect turning a turn a beat: the slowest frame in each 2 s went from 10-13.6 ms to 7.5-12.4 ms (a plain preset's reaches 12 ms too: the network stack on the same core), 100 frames a second throughout, no reset.
- **2026-11-17 A fast slider move no longer leaves an old value behind (firmware web pages, demo).** The live page sent a dragged slider's values (hue, saturation, brightness, colour temperature, effect speed, platter speed, preset, palette) every 120 ms and once more on release, each as its own request, and the unit answers requests side by side, so on a quick move away and back the last value (hue 0, say) could be applied before an earlier one, which then stuck: a hue shift left on the Art-Net / sACN picture that the slider no longer showed. Each control now has one request on its way at a time and keeps only the newest value to send next (`put()` in the live page), so the last value set is the last applied. The same goes for the live page's effect settings (`/control` `look`) and the settings page's minimum speed slider (`/setmin`). Reproduced in the demo with the unit's answers held back out of order: the old way ended at 54 degrees after a move to 0, the new one at 0.
- **2026-11-16 Tidying after the day's changes (firmware, tests, docs, demo, CI).** The editors (LEDitor and the live page's effect settings) fold the rows changed less often under **more**: drift, colours, churn, trail, punch and its rows, radial, steps, bounce and background; folded it says how many are set, the browser remembers it, and folded rows still count (`web_look.h`). The effect speed (EXP2, the live page, OSC `/fxspeed`) leaves a look locked to the platter alone, so it stays locked. On the local tempo (the tap, or after the show software stops) the beat clock runs on by itself at that tempo, four to the bar, from its last beat (`beat::localRun()`, keeping fractions of a millisecond so it doesn't drift), and each tap is a beat (`beat::localTap()`), so punch and the pulse hit on the taps; falling back to the tap keeps the clock instead of clearing it. Removed what nothing used any more: the beat flash's helpers (`presetmode` flash length and hold, `pulseFromUnit`, the intensity trim; `beat::flashDivider`, `flashOnBeat`; `ledfx::Flash`) and their tests; Art-Net's `Takeover` is `Signal` (`signalSeen`, `signalLive`). New automatic check that the browser demo draws what the unit does: `tests/parity_render.cpp` draws the looks in `tests/parity_looks.txt` (every effect and key) with the firmware's engine, `tests/parity.js` with the demo's, and the two are compared LED for LED (now at most 6 apart); both test runners run it, so CI does on every push. The user guide's screenshots are new, taken from the demo (`tools/capture_screenshots.py --demo`, with `demo.html?shot`): the live page, the LEDitor and the settings tabs, plus new Defaults, LEDs and Backup pictures; the Show and Colour ones are gone (the log and firmware update pictures still need a unit).
- **2026-11-16 The palette effect turns like the others; the platter from counted steps (firmware, tests, docs, demo).** Measured with every effect locked to the platter (speed 1), the palette effect was the odd one out: it turned the other way, at the speed divided by its spread (twice the platter at spread 128), because it scrolled the palette rather than turning the picture. With a spread above 0 the palette laid round the ring now turns as a whole, the place it starts included, at the speed and in the same direction as every effect (counter-clockwise for a positive speed), so with *Turns per platter* it stays with the platter; spread 0 is still the whole ring fading through the palette (preset 3 unchanged). Built-in presets 2 (Rainbow ring) and 12 (Forest drift) now turn the other way, and 12 at the speed it says (once round in 16 bars, not 8); their pictures are drawn again. With bounce, a spread palette rocks. The engine also follows the platter by the motor's own steps now: the step interrupt counts them (`platterSteps`, signed by `stepDir`, which loop() sets with the DIR pin) instead of the engine adding up the asked-for step rate, which the timer's whole-microsecond period misses by up to about 0.04% (a slip of about 4 degrees a minute at 33 RPM) and which led the timer during speed changes. The web pictures' platter edge follows the same count.
- **2026-11-16 Beat flash folded into punch; settings tabs rearranged; Art-Net takes the HSV (firmware, tests, docs, demo).** The global beat flash and each preset's punch did the same job, so the flash is now part of punch, per preset: `punchstyle=swell|flash` (swell as before; flash is full for its length then straight back, dark between at 255), `punchlen=2|4|8|16|32` (1/N of a beat, at least 12 ms on the ring at any tempo) and `punchevery=1-16`. Every (punch's, and the pulse's `every=`) takes any whole number up to 16 and simply counts beats on from switch-on, not tied to the bar: the effects get a beat count that never wraps (a double; the tempo tracker's own wraps at 64, which only powers of two divide). The editors show the three punch rows once Punch is above off. Gone with the flash: the live page's Beat flash row (on, every N, length, brightness), its settings, `/control` `flash`, `pulse`, `pulsediv`, `flashlen`, `/status` `pulseBri`, `pulseDiv`, `flashLen`, `flashOn`, the stored `flashon`, `pulsebri`, `pulsediv` and `flashlen` (cleared at the next Save & Reboot), and the bench-only LED mirror of the flash (`BENCH_LED_MIRROR`); OSC `/flash`, `/pulse` and `/pulsediv` are ignored with one log line. `/beat` still keeps the tempo on the beat. Settings: the Show tab is gone: its OSC listen port is on the Network tab and its starting BPM on the Colour tab, renamed **Defaults** (`#defaults`; links to `#show` and `#colour` open it). The Art-Net / sACN effect now takes the hue shift, saturation, brightness and colour temperature like every effect (at hue 0, saturation and brightness 100% and 6500 K it is as sent), so the LEDs tab's *Hue shift, saturation and brightness act on it* box is gone (`arthsv`, cleared at the next save). Backups made before still restore: the settings that are gone are read and ignored (`backup::kRetired`).
- **2026-11-16 Art-Net and E1.31 are an effect, not a takeover (firmware, tests, docs, demo).** Network input no longer takes the ring over when it arrives. It is an effect, **Art-Net / sACN** (`effect=artnet`, `FX_ARTNET`; built-in preset 23), picked like any other (the live page's Preset list and effect settings, the VAL knob, OSC `/preset` and `/look`, the LEDitor), with the crossfade on the way in and out. The LEDs tab's receive box now also decides whether it is offered: unticked, nothing arriving is used, and the effect and every preset using it are left out of the lists (`/names` gains `artnet` and `net`) and skipped by the VAL knob (`lightStepPreset()`, `ledPresetOffered()`); OSC `/preset` can still pick one. With nothing arriving it holds the last frame (dark before the first). It shows the colours as sent (or with the hue shift, saturation and brightness, with that LEDs-tab box ticked; never the colour temperature or flash), and like any pattern it turns by its speed per bar, beat or platter turn (with drift, bounce and steps): each LED takes the colour of the place that turns onto it, from the LEDs within 12 mm weighted by distance (`Geo` gains x and y in mm; on this ring no such place is over 11 mm from an LED, and neighbours are 12.7 mm apart, so speed 0 is exactly as sent); punch swells it on the beat. `/netmon` gains `shown`; the live page's Network box says whether the ring is showing it. The demo has a stand-in network picture (three coloured thirds round a white centre) so the turning can be seen.
- **2026-11-16 Counts and twist limited to what the ring can show (firmware, tests, docs, demo).** The ring is 178 LEDs about 7 rows deep, 13 round at its inner edge, 25 half way and 45 at the outer edge, and a repeating pattern needs a few LEDs per repeat (about 4 to look clean, under 2 it doesn't show). Each effect's count is now capped where it still shows (`ledfx::countMax()`): spiral arms 6 and its twist one turn (intensity 128, `kSpiralMaxTwist`; it went to two, which broke the curve into dots over 7 rows), segments 12 arms (3 bands with radial), sweep 6 heads (3 with radial), noise 6 repeats, ripples and pulse 3 rings; wipe colours and strobe flashes repeat over time and keep 24. A preset line asking for more still reads and draws at the cap, so nothing is rescaled (128 was already one turn of twist). The editors' sliders stop at the cap for the effect chosen (and follow radial), and twist reads in turns. Of the built-in presets only 11 (Fire spiral, 1.57 turns) was over, and is now 128 (one turn), as it draws; its picture is drawn again. Measured from the LED map: median spacing 12.7 mm.
- **2026-11-16 Backup and restore (firmware, tests, docs, demo).** A **Backup** page (`/backup`, in the menu's Settings list) downloads **Presets and palettes**, or **Everything** (every setting from the settings page too), as one text file in sections (`src/backupfile.h`, with host tests): `[network]`, `[settings]` and `[led]` hold the stored preference keys as key=value (the Wi-Fi password only if ticked, the admin password never), `[presets]` and `[palettes]` the two files' lines. **Restore** lists the parts a file has and restores the ones ticked, after checking all of it (a bad line refuses the lot, naming it): each part replaces the unit's as a whole (settings it doesn't name go back to factory; the Wi-Fi password stays if absent). Presets and palettes are in use at once and saved when the platter stops; network, settings and LED settings restart the unit. Network starts unticked, since it carries the unit's address. The demo has the page, with its presets and palettes.
- **2026-11-16 The web pictures stream (firmware, docs).** The live page's and the LEDitor's turntable pictures could stall for about 3 s under Art-Net: measured on the unit, the frames stopped while every request it got was answered in under 25 ms, so a frame request was lost on the network and the browser waited to try again. The pictures now read `/leditor/stream`, which sends frames (15 a second) one after another on one connection; it is sent from `oscTask` without ever waiting on a slow browser, a stalled stream is dropped and the page reconnects, and where there is no stream (the demo) the page polls as before. Every frame request now also gives up after 1.5 s instead of waiting for the browser. Drawing is cheaper too: each LED's glow is worked out once, so a frame (every LED lit, as under Art-Net) is only the adding up. The LEDitor's preview requests go out only on a change and once a second. Measured with timing logs (`src/rtdiag.h`, off) while Art-Net arrived at up to 104 packets a second: the ring rendered 100 frames a second throughout (longest gap 18 ms, longest frame 11 ms), so the LEDs were never affected; and with the platter turning the motor's step pulses came at exactly the rate set (1,779-1,781 pin changes in 2 s against 1,779) with the same spread round each (about 80 us either way) with and without the Art-Net, so the motor isn't affected either (its pulses come from a hardware timer on the other core).
- **2026-11-16 Live page: the whole preset editor in Effects (firmware, docs, demo).** The Effects box's **effect settings** now open by default and mirror the LEDitor's preset editor: every control including the look's own **Palette** (choosing it also sets the Palette override back to 0), and **Name** and **Save to** a preset number with **Save**, which replaces that preset (after asking, if it exists) through `/leditor/save` and then picks it. `/leditor/save` now takes lines up to `ledpreset::kLineMax` (it refused anything over 200 characters, which a line using every new key can pass).
- **2026-11-16 Publishing the documents (tools, CI).** Every push to `main` that changes `docs/` now publishes it to the web server (kennedy-williams.net/unclemagic/docs) over SFTP: `.github/workflows/publish-docs.yml` runs `tools/publish_docs.py --ci`, which uploads the files the push changed (`git diff` from the push's previous commit) and any the server lacks or has at another size; nothing on the server is deleted, and *Run workflow* fills in anything missing. The token is the repository secret `PUBLISH_TOKEN` (the host's SFTP token, renewed monthly on GitHub); the server, user and folder are the repository variables `PUBLISH_HOST`, `PUBLISH_USER` and `PUBLISH_REMOTE`. The server's host keys are pinned in `tools/publish_known_hosts` (ssh-keyscan); a key that doesn't match stops the upload. From a PC the same script (paramiko) uploads new, newer or resized files and stamps the server copy with the local time (`--dry-run` only lists, `--list` doesn't connect, `python docs/build_docs.py --publish` rebuilds first), reading its settings from the git-ignored `tools/publish.local.json` (template `tools/publish.example.json`) or `PUBLISH_*` environment variables.
- **2026-11-16 New effects, effect keys, crossfade, live looks (firmware, tests, docs, demo; host tests pass).** Three new effects: **ripples** (rings flowing out from the centre, or in), **sparkle** (LEDs lit at random, fading; the speed is how often each gets a chance) and **strobe** (the ring flashes *count* times a platter turn as the platter passes each place, so marks on the platter stand still or crawl; the guide warns about photosensitivity). New preset keys, each written only when not its default so older lines read back unchanged: `paint=auto|angle|radius|arm|tail` (how an effect picks its colours: a colour per arm, head or ring, or down a sweep's tail like a comet), `punch=0-255` (any effect swells on the beat: brighter and wider on it, dimmer between), `radial=1` (palette, wipe, segments and sweep run out from the centre), `steps=1` (the pattern moves only on the beat), `every=1|2|4|8|16` and `inward=1` (pulse) and `churn=0-255` (how fast noise changes shape). The sweep now takes `count` heads and the pulse `count` rings; every existing preset draws exactly as before (the preset pictures are unchanged). Three built-in presets: 20 Ocean ripples, 21 Party sparkle, 22 Platter strobe (on a unit with a saved list they appear after **Restore the built-in presets**, or by pasting their lines). Changing the preset or palette now **crossfades** over 500 ms. The live page's Effects box has **effect settings**: the running preset's own controls, the same as the LEDitor's preset editor (both now come from `web_look.h`), which run a *live look* until a preset is picked (**Back to the preset**); `/control?a=look`, and `/status` gains `look` and `lookLive`. OSC gains `/look <string>` (a preset line or its keys, applied to the running engine the same way) and `/fxspeed` (the effect speed, -16 to 16). The demo's engine and stand-in follow; a one-off comparison of the JS and C++ engines over 12 looks found at most 2 in 765 apart. New tests for each effect and key, the crossing-based strobe, the format's round trip with every key and the longest line (`kLineMax` 320).
- **2026-11-16 Live page: an Effects box (firmware, docs, demo).** The Light box was everything the ring does; it is two now. **Light** keeps the colour: hue shift, saturation, brightness, the user colour and the colour temperature. **Effects** (new, after it) has the preset and palette lists, a new **Effect speed** slider with a **1x** button, the white background and the beat flash (on, every *N* beats, length, brightness). The effect speed is the EXP2 knob's, until now not on the web: a quarter to four times the preset's own in the knob's steps (8 to a doubling), carried over between presets, until a restart. `/control?a=fxspeed&v=-16..16` sets it (`lightSetSpeed()`) and `/status` reports it as `fxSpeed`. The live page and the LEDitor remember the boxes' order in the browser; a box that order doesn't name (as Effects, for anyone who has moved a box) now goes after the box it follows on the page rather than at the end.
- **2026-11-16 Demo: the settings page (demo, docs).** `docs/demo.html` now has the settings page as well as Live and the LEDitor: the menu's Settings list opens each tab (with its heading), as on the unit, and `demo.html#settings/<tab>` opens it at one. The page is built in code on the unit, so `tools/build_demo.py` renders it from that code (`handleConfigRoot` in `web.cpp` and `ledAppendPanel` in `ledengine.cpp`): every piece of fixed text comes from the firmware, and the values the unit fills in are a unit as it comes (the LED settings and other defaults read from the sources, the built-in preset list, the stand-in's live figures). A value or condition the firmware adds that the demo doesn't know stops the build, naming it. The stand-in answers the Motor and fan tab's boxes that apply at once (`/setmin`, `/maxspeed`, `/fan`) as the firmware does; **Save & Reboot** shows a note and keeps nothing. The documentation index's demo card links to it.
- **2026-11-16 Settings page: each tab has a heading (firmware, docs).** With the tab buttons gone (the menu's Settings list opens the tabs), nothing on the page said which tab was showing; each tab now starts with its name as a heading (`h2.ph`), named as in the Settings list: Status, Network, Show, Colour, LEDs, Motor and fan, Password.
- **2026-11-16 Spirals follow the platter (firmware, tests, docs, demo; host tests pass).** Every spiral now turns the way the platter turns and curves back from that motion, whatever its unit: its speed is *with the platter* (positive) or *against it* (negative). Timed to the bar or beat, its rotation is multiplied by `Motion::follow`, which is the platter's direction eased: on a reversal it runs from one to the other over 5 s (`kFollowMs`; first 500 ms, slowed tenfold so the arms visibly wind round) through 0, so the spiral slows, stops and comes back the other way; locked to the platter it turned with it already. The twist's sign follows the same way (the outer ends lag the motion, like a whirlpool), so the arms straighten and bend the other way as the platter reverses. While the platter is stopped both are kept; before it has ever turned, the render task starts them from the way it is set to turn (`direction` and the speed's sign, as `/status`'s CW/CCW), and `follow` snaps to that first direction rather than reversing at switch-on. Built-in presets 7 (Rainbow spiral, with the platter) and 11 (Fire spiral, against it) change: under the usual clockwise platter both now turn the other way from before. The preset pictures' platter turns clockwise, as the unit's does, so presets 7, 10, 11 and 19 are drawn again. Other effects keep the speed's own direction. New tests for the follow, the curve and its mirror and straightening.
- **2026-11-16 Spiral trail that follows the platter (firmware, tests, docs, demo; host tests pass).** A new preset key, `trail=0-255` (the LEDitor's **Trail** slider, shown for the spiral): a glow behind each arm lasting that many hundredths of a second. The engine now measures how fast the pattern really turns (`Motion::rate`, turns a second, signed, eased over 150 ms; `advance()` takes the frame time), and the trail is *rate × glow time* long (in arm spacings, at most 0.95 of the gap), fading as the sweep's tail does, on the side behind the motion. So with `unit=platter` it follows the platter's speed, is gone while it stands still, and swaps sides through nothing when the platter reverses; with `unit=bar` or `beat` it follows the tempo. `Motion::platterDir` keeps which way the platter last turned, and a *sweep* or *wipe* locked to the platter now takes its tail side from it rather than the speed's sign alone (before, a platter-locked sweep's tail was in front of the head whenever the platter turned clockwise). trail=0 (every existing preset) draws as before; no built-in preset changes. The demo's engine and stand-in follow; new tests for the rate, the trail's side, length, standstill and limit, the platter-locked sweep, and `trail=` in the preset format.
- **2026-11-16 Hue, saturation and brightness on Art-Net input (firmware, docs, demo).** A new LED setting, **Hue shift, saturation and brightness act on it** (LEDs tab, Art-Net box; `ledCfg.artnetHsv`, prefs `arthsv`, off by default, saved by Save & Reboot). Ticked, the network's colours get the same hue shift and saturation as a preset (`ledfx::adjustColours`) and the eased brightness, from the HUE, SAT and VAL knobs, the live page and OSC; unticked (as before) they are shown as sent. The colour temperature and the beat flash never apply to network input, and the current limits always do. The live page's picture shows the result.
- **2026-11-16 The factory starting colour is white (firmware, docs, demo).** The user colour, which preset 1 shows at switch-on, starts white instead of red (`startRgb`, `userRgb` in `state.cpp`) until the Colour tab's **Starting colour** is saved. Every Save & Reboot stores the starting colour, so a unit whose settings page has been saved keeps the colour it had (red for units set up before this). White has no hue, so on preset 1 the HUE and SAT knobs change nothing until a colour is picked; the user guide and the FAQ (a new question) say so. The preset pictures in the guide (presets 1, 4, 8 and 10 use the user colour) are drawn in white again (`render_presets.cpp`), and the demo starts with white.
- **2026-11-16 FAQ rewritten for clarity (docs; the copy on the unit too).** Every answer now says what is happening, why, and what to do, in plain words rather than the guide's shorthand (*the network has the ring*, *as the unit comes*, *port-address*); multi-step answers (Art-Net, the beat flash, firmware updates) are numbered steps, and each question is its own heading, so it can be linked to. The setup Wi-Fi answer now says what the setup network is and why only the Wi-Fi page is on it (its password is the same on every unit, so nothing else is reachable there). Two answers corrected: the *applies at once* boxes are written to memory once the platter is stopped, and a firmware update must run 30 seconds with the network up.
- **2026-11-16 About page and the MIT License (firmware, docs, demo).** `LICENSE` at the repository root: the MIT License, copyright 2026 Steve Kennedy-Williams. The unit has an About page (`/about`, `ABOUT_HTML` in `web.cpp`), last in the menu's Help list: what the unit is, the firmware build and the unit's name and address (from `/status`, which gains `build`: the build date and time and the slot running), the copyright, a note that the firmware is built on the Arduino core for the ESP32 and other libraries under their own licences, and the licence text. The demo has it too (`demo.html#about`); the README and the user guide's menu table mention it.
- **2026-11-16 FAQ, a document of its own (firmware, docs).** A third document, `docs/FAQ.md` (built to `FAQ.html` and embedded like the other two), gathers the things that catch people out, from the user guide, as questions and short answers: phones and browsers (Android Chrome's *Nearby devices* permission, `.local` on Android, the `admin` user name, *offline*, per-browser box order, the pictures against the ring), lighting (Art-Net off as it comes and counted from 0, the network holding the ring, the VAL button, hue and saturation carrying over, the start-up state, the beat flash, a dim ring), saving (presets written once the platter stops, live-page changes lasting until a restart), tempo, and network, updates and power. The unit serves it at `/docs/FAQ.html`, and the menu's Help list has **FAQ** between Hardware overview and Log; the guide's Troubleshooting points to it, and the documentation index and README list it. The connection log in `poolserver.h` (`WEB_TRACE`) is off again, kept for next time.
- **2026-11-16 Dragging boxes on a phone; Android's local network permission (firmware, docs, demo).** On a touch screen (`pointer:coarse`) the grip (⋮⋮) of the live page's and the LEDitor's boxes is a finger-sized target (about 58 × 58 px with a bigger symbol, was about 20 px), and while a box is dragged the page scrolls when the finger is within 80 px of the top or bottom, so a box can go past the screen's edge. Android Chrome's *ERR_TOO_MANY_RETRIES* with the unit's address turned out to be the phone, not the unit: newer Android lets an app onto the home network only with its **Nearby devices** (*Local network*) permission, and Chrome's requests never left the phone (the unit's connection log showed none) while Firefox, allowed, worked. The user guide's *Finding it* says so. The turntable picture (live page and LEDitor, `web_ring.h`) no longer draws each LED's glow as its own gradient fill blended onto the canvas, 178 a frame at 10 frames a second, which left Firefox on a phone barely responding; the glows are summed in script on a layer a third of the size and laid on the canvas in one step, scaled up smoothly (the same picture: on average 0.5 in 255 apart, at most 12).
- **2026-11-16 Web server no longer waits on idle connections: Android Chrome reaches the unit (firmware).** The core's WebServer takes one connection at a time and waits up to 5 s for it to send its request; browsers, Android Chrome above all, open spare connections in advance and leave them idle, so each one held every real request up by 5 s. Measured on the unit: 1 idle connection made a page wait 4.8 s, 2 made it 9.8 s, 4 made it 19.8 s, and with the password prompt and the page's own requests on top, the phone gave up. `configServer` is now a `PoolServer` (`src/poolserver.h`, a subclass with its own `handleClient()`): waiting connections go into a pool of 8 and whichever has its request in is served, oldest first; idle ones are dropped after 15 s, or the oldest when the pool is full. Reading and answering a request is the core's own code, so routes, the password and the clean close after each reply are as before.
- **2026-11-15 Network input monitor (firmware, docs, demo).** Reviewing why Resolume's Lumiverse didn't reach the ring: Art-Net (UDP 6454) and E1.31 (UDP 5568) are received on every link, but receiving is **off** as the unit comes (LEDs tab), only ArtDmx/E1.31 data for the first universe (and the next) is used, and that first universe is Art-Net's 15-bit port-address counted from 0, where some programs count from 1; anything else was dropped without a trace. Now the sockets are open whenever the unit has an address, receiving on or off (with it off nothing is applied), and every DMX packet is noted by protocol and universe with its last sender, count, age and whether it was used; other packets on the ports (ArtPoll, ArtSync, ...) are counted with their OpCode. `GET /netmon` reports it, and the live page has a fourth pill box, **Network** (folded at first, polling once a second while open), that shows it with the first universe's Net/SubNet/Universe. A box added since a browser kept its order (as Network is) now goes at the end of that order rather than the top. The LEDs tab's Art-Net tip explains the numbering; its preset list tip no longer calls palette 1 *unit* or the LEDitor the simulator page.
- **2026-11-15 LEDitor picture: the presets and palettes as written (firmware, docs, demo).** The LEDitor's Turntable picture leaves out the hue shift, saturation and brightness (and the current limits), so it shows a preset or palette exactly as written while it is edited; the colour temperature still shows. The engine keeps a copy of each frame from before the HUE and SAT controls for it (`rawPx`; `/leditor/frame?full=1`). The LEDitor's **Full brightness** box and `/leditor#full` are gone, as the picture is always at full brightness. The live page's picture is unchanged: what the ring gets, before the limits.
- **2026-11-15 Web server: a clean close after every reply (firmware).** Checked against a hand-written server that Android Chrome is happy with: the ESP32 core's WebServer already reads the request before replying and sends HTTP/1.1 headers with CRLF line ends, a Content-Length (or chunked) and `Connection: close`, and our pages only reply through it. What it doesn't do is end the connection gently: it closes the socket as soon as the handler returns, and a socket closed with unread bytes waiting is aborted with a TCP reset, which can cut a reply short. A middleware (`webBegin()`) now runs after every handler: it reads and drops anything left of the request, shuts down the sending side (so the FIN follows the reply) and waits 2 ms before the core closes the socket. The core's 5 s wait for an idle connection's request (`HTTP_MAX_DATA_WAIT`, which Chrome's unused pre-opened connections can trigger) is unchanged: it can't be set from the sketch.
- **2026-11-15 Web pictures before the current limits (firmware, docs).** The turntable picture on the live page and the LEDitor (`/leditor/frame`) is now taken after the brightness and colour temperature but before the per-LED and ring current limits; the ring itself is still held to them. So at 100% the picture matches Full brightness, as in the demo, and white can look brighter on the screen than on the ring.
- **2026-11-15 Demo: no current limiting (demo).** The demo's stand-in unit no longer applies the per-LED and ring current limits, so its picture (the live page's Turntable and the LEDitor) shows the pattern at the brightness set, white included: at 100% it is the same as Full brightness. The unit itself keeps its limits.
- **2026-11-15 Bench units ignore the ring's temperature sensor (firmware, docs).** If the TMC2209 doesn't answer at start-up, the unit is taken to be a board on the bench with nothing attached (`benchMode`): the LED ring's temperature sensor is still read and shown but limits nothing, so the ring keeps its set limit and the fan follows the chip alone. On the bench the sensor input read shorted (146 C, then shorted), which held the ring at the lowest limit, 1500 mA, with the fan full, so everything bright looked dim. It is logged at start-up (*BENCH: ...*), reported in `/status` (`bench`), marked on the Status and LEDs tabs and on the live page's bottom line. A unit whose motor driver answers always uses the sensor.
- **2026-11-15 Live page: no Full brightness box (firmware, docs, demo).** The live page's Turntable picture is always the frame as sent to the ring (brightness, colour temperature and the current limits included); **Full brightness** is the LEDitor's alone, for previewing while editing.
- **2026-11-15 Full brightness picture shows the colour temperature (firmware, docs, demo).** The turntable picture's **Full brightness** (live page and LEDitor, `/leditor/frame?full=1`) was the frame before the brightness, the colour temperature and the limits, so moving the colour temperature changed nothing in it. It now has the colour temperature applied (a colour, not a brightness; the engine multiplies the rendered frame by the same gains), and still leaves out only the brightness and the limits. The frame as sent was already right.
- **2026-11-15 LEDitor in foldable, draggable boxes; a white background on the live page (firmware, docs, demo).** The LEDitor is three boxes like the live page, **Turntable** (the picture, with Full brightness), **Preset editor** and **Palette editor**, each folding away by a purple pill and draggable into any order by a grip (⋮⋮), both remembered in the browser. The fold and drag code (`fold()`, `boxes(key)`) and its styles moved into `src/web_box.h` (`BOX_JS`, `BOX_CSS`), shared by both pages. The LEDitor's status line (*live, 12 frames/s · LED settings are on the LEDs tab*) is gone: the menu bar's address reads *offline* when frames stop and comes back when they resume (`window.umOk()`), and the LEDs-tab link is in the Turntable tip. The **Soft** box is gone; the picture is always drawn soft. The live page's Light box gains **Background: white instead of off**, the preset editor's background for whatever preset is running (`lightWhite`, `/control?a=white`, `/status` `white`; unticked, each preset's own; until a restart).
- **2026-11-15 Fewer headings: the menu's Settings list is the settings page's tabs (firmware, docs).** The settings page loses its heading and its row of tab buttons: the menu's **Settings** list opens each tab (`/#status` .. `/#security`, as before), marks the one showing, and the page still follows the address, so bookmarks and older links work. The LEDitor page loses its heading too, and the Settings list its subheadings (*Settings* and *The unit*: Firmware update and Wi-Fi setup simply follow the tabs).
- **2026-11-15 Beat flash length as a fraction of a beat (firmware, tests, docs, demo; host tests pass).** Each flash lasted a fixed 80 ms; it now lasts 1/32, 1/16, 1/8, 1/4 or 1/2 of a beat (1/8 as it comes, 62 ms at 120 BPM, close to the old 80 ms), chosen in a list beside **Every N beats** on the live page and kept across restarts the same way (`/control?a=flashlen&v=..&save=1`, `flashLen`/`startFlashLen`, prefs `flashlen`, written once the motor is stopped by `saveLenIfIdle()`; `/status` reports `flashLen`). `presetmode::flashHoldFor()` works it out from the tempo, never under 12 ms so a frame at 100 a second always shows it (1/32 at 300 BPM would be 6 ms). The bench LED mirror keeps its fixed 80 ms. New tests for the hold times.
- **2026-11-15 One-line menu with Settings and Help lists; live page as three pill boxes (firmware, docs, demo).** The menu bar on every page (`web_nav.h`) is **Live | LEDitor | Settings ▾ | Help ▾** and, at the right, the unit's address and OSC port (from `/status` once the page loads; `window.umIp()`; empty where there is no `/status`). **Settings** lists the settings tabs, then Firmware update and Wi-Fi setup; **Help** the User guide, the Hardware overview and the Log. The Menu button and the brand name are gone; a list is marked when one of its pages is open, opening one closes the other. It fits on one line at 375 px. The live page loses its heading row (the address is on the menu bar, and reads *offline* there when the page loses the unit) and is three boxes with purple fold pills: **Turntable** (folded at first; the picture streams only while it is open, with Full brightness inside), **Light**, and **Tempo** (was *Tempo and platter*; its **more**/**less** pill shows while it is open). The boxes can be dragged into any order by a grip (⋮⋮) in their titles (pointer events, so touch works). Each fold and the order are remembered in the browser.
- **2026-11-15 Live page heading: the address instead of a status line (firmware, docs, demo).** The heading is **Live** followed by the unit's address and OSC port (`ip:3335`, Ethernet's address, else Wi-Fi's), and the line under it (*live · liquidturntable.local*) is gone; when the page loses the unit the address reads *offline* in orange. **Show turntable** (shortened from *Show the turntable*) sits on the heading line, which fits at 375 px wide with an address like `192.168.1.50`; **Full brightness**, once shown, wraps under it. `/status` gains `oscPort`. The fold pills (**Light**, **more**/**less**) are filled with the demo bar's purple (`#6a5acd`) at 80% opacity, with light text, instead of outlined.
- **2026-11-15 Palette 1 is the user colour, 2 red/blue/white; Live page: hue, saturation and brightness first (firmware, tests, docs, demo; host tests pass).** The two swap places, so the user colour comes first in every list (`PAL_UNIT` 1, `PAL_REDBLUEWHITE` 2; the built-in, editable palettes are now 2-11). Preset lines and `/ledpalettes.txt` name their palettes, so saved lists are unaffected; only the numbers for the EXP1 knob, OSC `/palette` and `/control?a=palette` change (and a hand-typed `palette=1` now means the user colour). The LEDitor's lists follow. On the live page, **Hue shift**, **Saturation** and **Brightness** now come before the **Preset** and **Palette** lists. The **Light** box now comes first and folds away (a pill beside its title), and **Tempo and platter** opens in a short view: one line with the BPM, the platter's RPM and the speed setting with its direction, the not-in-tempo warning, then the speed slider and **Tap** on one line; **more** beside its title shows the rest (where the tempo comes from, Start/Stop, Slower/Faster/Reverse, Set BPM, the speed limit). The browser remembers both.
- **2026-11-15 No colour mode: hue and saturation act on every preset; new knob jobs; Live page lists (firmware, tests, docs, demo; host tests pass).** The ring always runs a preset. **HUE** and **SAT** are now a hue shift (every colour turned round the colour wheel; 0 = as drawn) and a saturation scale (1 = as drawn, 0 = white) that the engine applies to every frame a preset draws (`ledfx::adjustColours()`; greys, white and black are left alone, and 0 and 1 leave the frame exactly as it was). **VAL** is still the brightness, but the **VAL knob** picks the preset until a press of its button switches it to brightness (and back; `valKnobPreset`, `/status` `valKnob`). **EXP1** picks the palette (its button: 0) and **EXP2** the effect speed (its button: the preset's own); the effect-intensity knob is gone. The HUE button sets the hue shift back to 0; the SAT button still toggles white. Palette 2 is renamed **usercolour** (*user colour*; `unit` still reads): one colour, `userRgb`, set by the live page's colour picker (`/control?a=colour`), starting as the Colour tab's starting colour, and preset 1 is *User colour*, solid in it, so with the user colour red (as it comes) the HUE and SAT knobs give exactly what colour mode did. The unit starts on preset 1. The **live page** loses the Colour/Preset buttons; preset and palette are lists (*N: name*, refreshed when the tops change), with the user colour picker and Hue shift (degrees), Saturation and Brightness sliders. `/status` drops `lightMode` and adds `userColour`. OSC `/presetmode 0` (and `/control?a=mode&v=0`) picks preset 1, `1` does nothing. The LEDitor's palette list shows *user colour*. New tests: hue shift and saturation on colours and frames, the palette's name and old alias, an old `palette=unit` line read and written back as `usercolour`. The live page's screenshot still shows the earlier layout.
- **2026-11-14 Demo bar: its description behind an (i) tip (demo).** The demo's top bar is now just **Demo**, a round (i) and **Start over**; the text about the stand-in unit opens from the (i), styled like the unit's own tips, and closes on a click anywhere else, in the bar or in the page under it.
- **2026-11-14 Speed setting goes to 200% (firmware, tests, docs, demo; host tests pass).** `speedScale` runs from -2 to 2 (-200% to 200%) instead of -1 to 1, through one constant, `speedScaleMax`: the speed knob (and OSC `/scaleEncoder`), **Faster**, `/control?a=speed`, the live page's **Speed** slider and the saved home speed. 100% is still in time with the tempo and **Start** still sets it; above 100% the platter runs faster than the beat (200% is twice the tempo's speed), and the speed limit caps it as before, with the not-in-tempo warning when it does. `motion::step()` already scaled past 1. New test: 60 BPM at 200% runs at twice the speed uncapped, 120 BPM at 200% is held at the limit and reported; the fuzz covers -2 to 2.
- **2026-11-14 The simulator is called the LEDitor (firmware, docs, demo).** The menu bar's *Simulator* is now **LEDitor**, and the page's title and heading, the menu entry (*LEDitor: ring picture, preset and palette editors*), the live page's tips, the demo and the documents follow. Its address is now `/leditor` (`/leditor#full` ticks Full brightness), and its data requests moved with it (`/leditor/map`, `/leditor/frame`, `/leditor/preset`, `/leditor/save`, `/leditor/preview`, `/leditor/palettes`, `/leditor/palette`, `/leditor/palsave`, `/leditor/palreset`; the live page's turntable picture uses the new ones). The old `/sim` answers with a redirect to `/leditor`, so bookmarks keep working (`#full` included); the old data paths are gone. The User Guide's section is *The LEDitor* (`#the-leditor`), and the demo opens at it with `demo.html#leditor` (`#sim` still works).
- **2026-11-14 Flash on every Nth beat moves to the live page (firmware, docs, demo).** The live page's **Bar flash** switch is replaced by the Show tab's *Flash on every Nth beat* field, now beside **Beat flash** as **Every _N_ beats** (1-16, greyed out while the flash is off). The Show tab no longer has it, so the live page is where it is set and it is kept across restarts from there: `/control?a=pulsediv&v=..&save=1` sets `startPulseDiv` too, written by `loop()` once the motor is stopped (`saveDivIfIdle()`, prefs `pulsediv` as before). The settings form leaves it out and keeps the saved value; OSC `/pulsediv` still lasts until a restart.
- **2026-11-14 Speed limit in RPM, applied at once; Bar flash switch (firmware, docs, demo).** The Motor & fan tab's *Max speed* (steps/s, in the Save & Reboot form) is now a **Speed limit** box in RPM (1-187.5) marked *applies at once*: **Apply** posts the new `/maxspeed?rpm=..&save=1`, which sets the live cap (`maxSpeed`, read by `loop()` every pass) and the kept one (`startMaxSpeed`, prefs `maxspeed` in steps/s as before), written once the motor is stopped (`saveMaxIfIdle()`). A form without the field keeps the saved limit. The live page shows the limit under the BPM row with a box and **Set limit** (`/control?a=maxrpm`, until the next restart); `/status` gains `maxRpm`, and the not-in-tempo warning names it. Beside **Beat flash**, a **Bar flash** switch sets the flash to the first beat of each bar (`/control?a=pulsediv` 4, or 1 when unticked), greyed out while the beat flash is off.
- **2026-11-14 Speed cap applies after the speed setting, not to the tempo (firmware, tests, docs, demo; host tests pass).** `recalcDefaultSpeed()` no longer clamps the tempo's speed to `maxSpeed`; `motion::step()` caps the platter's speed after `speedScale` (as it already clamped its output). So at 300 BPM and 50% the platter turns at 12.5 RPM, half the tempo's speed and in time, where it used to run at half the cap (8.5 RPM). At 100% nothing changes. The clamp's old job, keeping the BPM ramp from winding up past what can be driven, moves into `step()`: the ramp is held at the tempo speed that reaches the cap at the current `|speedScale|` (`motion::rampForCap()`), which drives the same speed, so a lower tempo after running at the cap still shows at once. `/status` `tempoCapped` now means the cap is holding the platter back at the current speed setting (`motion::capped()`); the live page's warning says so. New tests: 300 BPM at 50% runs uncapped, at 75% at the cap, back to 50% glides; no wind-up from 500 BPM; the fuzz now covers tempos up to 500 BPM.
- **2026-11-13 Live page: Full brightness for the turntable picture (firmware, docs, demo).** Beside **Show the turntable**, shown while the picture is, a **Full brightness** box draws the frame before the brightness, the colour temperature and the current limits (`/sim/frame?full=1`), as the simulator's does; kept in the browser like the other box.
- **2026-11-13 Docs on a web server: pages are no longer cached for a day (docs).** The project's site sends every file with a one-day cache time, and a cache in front of it kept serving the demo from before the live page's new controls to every browser (phones included, whatever their own cache) although the new file was on the server. New `docs/.htaccess` (Apache only) makes HTML and Markdown pages revalidate on every visit (`Cache-Control: no-cache`), so an update shows at once while unchanged pages still come from the cache. Images keep their cache time; the documents already link them with `?v=<date>`.
- **2026-11-13 Live page: the turntable picture, a speed slider and a colour picker (firmware, docs, demo).** **Show the turntable** (remembered in the browser, off at first) draws the simulator's ring picture, with the turning platter edge, on the live page from `/sim/frame` about 10 times a second while the page is in view. A **Speed** slider sets `speedScale` from -100 to 100% (0 stops, below 0 the other way) through the new `/control?a=speed`. In colour mode a **Colour** picker sets hue and saturation together through the new `/control?a=colour&v=rrggbb` (`colour::rgbToHsv`), leaving brightness on its slider. The ring drawing moved to `src/web_ring.h` (`RING_JS`), which the simulator and the live page share; `/control` values are now URL-encoded by the page. The demo's stand-in answers both new actions. `speedScale`'s default in `state.cpp` is now 0 (it was 1.0, though `setup()` already set 0 at switch-on, so the platter always started stopped).
- **2026-11-13 Simulator: an edge that turns with the platter (firmware, docs, demo).** The simulator's picture has the platter's edge round it, with a tick every 10 degrees, longer ones every 90 and a violet mark, turned to the platter's angle. The LED engine adds up the platter's turns every frame from the motor's step rate (the same turns the platter-unit effects use) and `/sim/frame` sends the angle as two more bytes (536 in all: a little-endian 16-bit fraction of a turn, counter-clockwise from above). There is no position sensor, so the violet mark is where the platter was at switch-on. The demo's stand-in sends the angle too.
- **2026-11-13 Beat flash is a setting, off by default (firmware, docs, demo).** Until now every `/beat` from the chosen OSC beat sender flashed the ring, and the only way to stop it was a flash brightness of 0. The Show tab's Beat flash box has a **Flash on the beat** switch (`flashon` in prefs, default **off**, so a unit updated to this firmware stops flashing until it is switched on); the live page has a **Beat flash** switch (greying the brightness slider while off), `/control?a=flash&v=0|1` and OSC `/flash 0|1` change it until a restart, and `/status` reports `flashOn`. Off, `/beat` still keeps the beat clock. In the demo the stand-in unit no longer flashes on every tempo beat (it did, at a flash brightness below VAL, which showed as dark blinks after going from Live to Simulator); with the live page's switch on it flashes on its tempo beats in place of an OSC beat sender, and its flash brightness starts at the firmware's 255.
- **2026-11-13 Live and Simulator demo in one HTML file (docs, tools; checked in Chrome over http and from disk).** `docs/demo.html` runs the unit's own Live control and LED simulator pages with no turntable, linked from the docs index (`demo.html#sim` opens at the simulator). `tools/build_demo.py` (`--check` to verify) takes the two pages from `src/web.cpp` and `src/ledsim.cpp` with the theme and menu macros put in, and the built-in presets, palettes and LED positions from `ledpreset.h`, `ledfx.h` and `led_map.h`. Each page gets a few lines at the top that send its requests to a stand-in for the unit (`tools/demo/unit.js`: `/status`, `/control`, `/names` and every `/sim/...` endpoint), which runs a JavaScript port of the LED engine (`tools/demo/engine.js`, from `ledfx.h` and `palstore`'s spread) 100 times a second, with tempo, the beat flash, platter turns, brightness, colour temperature and a simplified current limiter. The menu's Live and Simulator links switch pages; the other pages say they are only on the unit. Presets and palettes saved in the demo last until a reload; **Start over** also clears the editors' drafts. `engine.js` and `unit.js` are hand ports, so a change to the engine or the endpoints needs them changed as well.
- **2026-11-13 CPL rotation scripts keep what they can't fetch (tools).** Both `check_cpl_rotation.py` copies (ring and mainboard) now keep the last run's turn (and, on the mainboard, the origin move) for any part whose EasyEDA footprint doesn't come back, and stop without writing anything when none does. A run against a throttled EasyEDA had written a mainboard rotation file 37 entries short (restored from git before use). Tested offline: with the LEDs' fetch failing, all 178 kept their 180. The docstrings say to pass the easyeda2kicad interpreter by its full path.
- **2026-11-13 LED ring: connector corner moved for the heat plate, CP1 surface-mount (board; DRC and ERC clean, not built).** J1 moved 4 mm up and 1 mm left, and J2, R1 and R2 4 mm up with it, tracks and vias following, so the heat plate's window round J1's pins is a closed hole with 2 mm of clearance instead of a notch to the edge. CP1 is now the SMD ROQANG RVT1C102M1010 (1000 uF 16 V, D10 x L10.2, C72470) on CP_Elec_10x10, with three ground vias; J2 stays through-hole (no surface-mount 5-pin JST XH is stocked at LCSC). 360 SMD + 2 THT parts. Tools: easyeda2kicad installed, and `check_cpl_rotation.py` (now given the interpreter's full path) decided every part, CP1 +0; it no longer overwrites the offsets file when nothing could be fetched. `gen_led_map.py` used the old SK6812 data pins (2, 4), so its `--check` had failed since the SK9822 change; it now uses SDI (1) and SDO (6), and `led_map.h` is unchanged. Gerbers, drill, BOM/CPL, PDFs, cost and reports regenerated; the hardware overview updated.
- **2026-11-13 Enclosure: inset aluminium backplate, ring heat plate, earthing (enclosure and docs; drawings not cut or fitted).** The rear of the box is now a 343 x 74.9 mm, 2 mm 5052-H32 aluminium panel in place of the plywood rear wall, set inside the box for leak protection: it sits in 2.4 x 3 mm grooves in the floor, the side walls and the lid underside, with a 4 mm plywood lip behind it, drops in from above with the lid off, and is also screwed into the two rear corner blocks (the grooves alone leave some slop); the lid overhangs it as a drip edge. The interior is unchanged; the outside is 349 x 284.4 x 80.9 mm. `backplate.dxf` (from `layout.scad`, `part="backplate2d"`, seen from outside) has the IEC inlet with switch and fuse (short side still a placeholder), the etherCON and USB D-size holes (24 mm, with screw holes on both diagonals of the 19 x 24 mm pattern), an M4 earth stud, four corner-block screw holes and 114 vent holes. The LED tray is now the ring's heat plate: 1/8 in 6061-T6, black anodised, on a 0.5 mm thermal pad, held by the four M3 screws on the ring's real 183.84 mm hole square (the model had 187); `ring_plate.dxf` and `tinkercad/ring_plate.svg` come from the ring board by the new `tools/gen_ring_plate.py`, with one closed window round the through-hole pins of J1 and J2 (2 mm all round, 1.87 mm of metal round the corner screw, 3.1 mm to the board edge). The model imports the DXF; the tray is 0.7 mm thicker, so the plate-to-tray gap is 2.9 mm. `MOTOR_MOUNT.md` gains "Ring heat plate" and "Earthing": both plates earthed at the backplate's stud, the low-voltage ground left off both. STLs, SVGs and renders regenerated.
- **2026-11-12 Palette 1 is red/blue/white; new built-in palette ember (11); your own palettes are 12-19 (firmware, docs; host tests pass).** The *colours* palette (a loop through each preset's `c1`-`c3`) is gone: the palette editor can now edit any palette, so palette 1 is a built-in, editable `redbluewhite` (red, blue, white: what *colours* defaulted to), and a new built-in `ember` (dark red, orange, red) takes number 11, which moves user1-user8 from 11-18 to 12-19 (saved lines and `/ledpalettes.txt` name them, so only knob and OSC numbers for them move). Preset 5 uses ember and preset 9 redbluewhite; their pictures are byte-for-byte the same as before. Older preset lines still load: `palette=colours` (or `colors` or `1`) reads as redbluewhite, `c1`-`c3` are still checked but unused, and *colours* with exactly ember's colours (the old preset 5) reads as ember, so a list already on a unit keeps its look; lines are written without `c1`-`c3`. The preset editor loses its colour pickers; the palette lists, the palette sheet and the docs show the ten built-in palettes. An empty user slot is drawn as redbluewhite. `palstore.h` now keeps its tables by palette number.
- **2026-11-12 Simulator: `/sim#full`, narrower colour pickers (firmware and docs).** The simulator opens with **Full brightness** ticked when the address ends in `#full`, and ticking or unticking it (or turning Preview on, which ticks it) keeps the address in step, so it can be bookmarked. The screenshot script takes `/sim#full`, so the guide's picture shows the pattern rather than the limited frame. The preset editor's three colour pickers are narrower and stay on one line at phone width.
- **2026-11-12 Simulator screenshot (docs, tools).** `tools/capture_screenshots.py` also takes the simulator page (`webui_sim.png`: the ring picture and both editors; pages up to 4000 px tall). The User Guide's simulator section and the docs index show it.
- **2026-11-12 Preset list as "N: name" (firmware and docs).** The simulator's preset editor picks the preset number from a dropdown of all 64 slots, each shown as `N: name`, `N: (no name)` or `N: (empty)` from `/names`, refreshed after each save; choosing one only sets the number to save to, and **Load** reads it in. The live page's preset and palette labels read `N: name` too.
- **2026-11-12 Screenshot script for the User Guide (tools; tested against a local test server, not yet on the unit).** `tools/capture_screenshots.py` takes the nine web UI pictures in `docs/screenshots/` (every settings tab, the live page, the log, firmware update) from a running unit with headless Chrome or Edge at 480 px, and crops the empty background off the bottom. It asks for the web password (or reads `TURNTABLE_PASSWORD`) and keeps it in the process. Chrome won't let a page's own requests carry a password written into the address, so the script runs a small local proxy that adds the sign-in to each request; it passes only GET requests, so nothing on the unit can change while the pictures are taken. `--host`, `--only` and `--out` pick the unit, the pictures and the folder. Stdlib and Pillow only.
- **2026-11-12 Web pages tidied: usage notes behind (i) tips, clearer tabs and boxes (firmware and docs; not checked on the unit).** A shared `TIP()` (`web_theme.h`) puts each usage note behind a round **i** beside a box's title or a setting's name; it is a CSS-only `<details>`, its note spans the box, and the menu's script closes it on any other click. Notes moved into tips on the settings page, the LEDs tab, the live page, the simulator's two editors and picture, the firmware update page and the Wi-Fi page; state and warnings stay visible. Settings tabs: **Show** now holds OSC, **Starting BPM** (from Motor) and one **Beat flash** box (brightness and the every-Nth-beat divider, which were two boxes on the Colour tab); **Motor** is **Motor & fan** (Motor, Slow-speed floor and Cooling fan, the last two badged *applies at once*); **Security** is **Password**. The Status tab and the LEDs tab's engine box are tables instead of line-broken text. LEDs tab boxes renamed (LED engine now, LED type and wiring, Current limits, Ring temperature, Art-Net / E1.31 input, Preset list, with a link to the editors). The live page is grouped into **Tempo and platter** and **Light**; the simulator's Full brightness and Soft sit under the picture. Field names and tab ids are unchanged, so saved settings and `/#tab` links still work. The User Guide's screenshots show the earlier layout.
- **2026-11-12 Palette editor on the simulator page; both editors kept across a visit to the live page (firmware and docs; host tests pass).** The eight named palettes (rainbow .. heat, 3-10) can be edited: 1-16 colours each, changed with a swatch or a hex value, inserted (+, half way to the next), deleted, added at the end and reordered by dragging a handle (pointer events, so touch works). A strip shows the palette as the effects use it. **Save palette** puts it in use at once for every preset, the SAT knob and OSC; **Save as new** puts the colours in the first empty one of eight user slots instead (`user1`..`user8`, palettes 11-18, with an optional name of up to 23 characters, shown in both editors and on the live page), leaving the palette it started from as it was. **Reset to factory** (this palette) and **Reset the built-in palettes** put the FastLED colours back and keep the user palettes; for a user palette the button is **Delete**. The SAT knob, `/palette` and the live slider go up to the last filled slot (`ledPaletteTop()`), and a preset naming an empty slot shows its own colours. With the preset editor's Preview on, unsaved palette changes show on the ring (`pal=` on the frame requests); **Use it in the preset editor** points the preset at the palette. New `src/palstore.h` (PC-tested, `tests/test_palstore.cpp`): stops, their spread over the engine's 16 entries (16 stops = the factory tables exactly), and `/ledpalettes.txt`, which holds only changed palettes and is written by `ledService()` when the platter stops. New endpoints `/sim/palettes`, `/sim/palette`, `/sim/palsave`, `/sim/palreset`. The preset and palette editors keep their contents (unsaved changes too) in the browser's localStorage, and Preview's on/off for the tab, so going to Live and back finds them as they were.
- **2026-11-12 Drawing of the one-universe Art-Net pairs (docs).** `docs/diagrams/artnet-single-pairs.svg` shows the ring from the front with the eight LED pairs that share an element in the **One universe** layout, and the element and DMX channels each pair takes (D1+D2 ch 1-3 ... D177+D178 ch 508-510). It is in the User Guide's Art-Net section and published in `docs/files/diagrams/`. Generated by `tools/gen_artnet_pairs.py` from `src/led_map.h` and `artnet::kSinglePairFirst`, so it follows the firmware; `--check` reports a stale drawing. Colours follow the page theme. Hovering over any of the 178 LEDs shows a label (CSS only, no script; no browser tooltip) with its one-universe element and channels, the LED it shares with if any, and its address in the default two-universe layout.
- **2026-11-12 Expansion knobs: EXP1 effect speed, EXP2 effect intensity; Art-Net in one universe (firmware and docs; host tests pass, not tested with knobs or Art-Net).** In preset mode EXP1 scales the preset's speed, 8 clicks a doubling from 1/4x to 4x, and EXP2 moves its intensity 4 a click; their buttons put each back to the preset's own. Both are offsets held across preset changes (`lightSpeedClicks`, `lightIntensityTrim`; arithmetic in `presetmode.h`), applied in the render task after the preset is resolved, so the editor preview is unaffected. EXP3 is still only logged. New LEDs-tab setting **One universe** (`art1`): the ring answers on the first universe only, as 170 elements (510 channels). Eight pairs of chain-adjacent LEDs on the outer edge share an element: D1+D2, D3+D4, D5+D6 (top arc), D83+D84 and D95+D96 (the sides), D173+D174, D175+D176, D177+D178 (bottom arc), each 12.7-14.2 mm apart; every other LED keeps its own, in chain order. Software only: the ring still drives all 178 LEDs. `tests/test_artnet.cpp` checks the pairs against `led_map.h`.
- **2026-11-11 Bounce: out-and-back motion for the LED effects, heartbeat as pulse with bounce (firmware and docs; host tests pass, not tested on a ring).** New preset key `bounce=1` (written only when on) and a **Bounce** box in the preset editor. The motion runs one turn forward at the preset's speed and unit, then the same turn back (`Motion::swing`, `back`, `cycles` in `ledfx.h`, advanced alongside the plain rotation). Sweep swings like a pendulum with its tail trailing both ways; segments, spiral and noise rock; palette and solid run to the palette's last entry and back (index 240, so a palette that doesn't loop never crosses its seam); wipe lays colour 1, 2, ... of its count over colour 0 and takes it back off. Pulse with bounce is the heartbeat: it leaves the centre on a beat, reaches the edge on the next and is back on the one after, eased at each end. `effect=heartbeat` reads as `effect=pulse bounce=1`. Effect numbers are unchanged.
- **2026-11-11 Ring temperature log: no false "back to the set limit" on a sensor fault (firmware; host tests pass).** The derating log line followed the sensor state as well as the limit, so a hot reading followed by a shorted or open sensor logged "LED ring at nan C: ring limit back to 4500 mA" although the limit stayed at the fault value (1500 mA shorted, 3000 mA open). The limit itself was right. `ringtemp::derateLog()` now decides from the limit alone, and the fault keeps its own warning; `tests/test_ringtemp.cpp` runs the bench sequence.
- **2026-11-11 Expander read every 2 ms; expansion encoders decoded (firmware; host tests pass, ESP32 compile checked, not tested on hardware).** The TCA9555 is now read by its own task (`expanderTask()` in `panel.cpp`, core 1, priority 2, above `loop()`) every 2 ms instead of from `loop()` every 20 ms, and the I2C bus runs at 400 kHz (was the 100 kHz default), so a read takes about 0.15 ms. The task is the bus's only user after `setup()`; it publishes each sample in one 16-bit store, decodes `J-ENC-EXP1`-`3` with the same quadrature decoder as the native encoders (`encExp1`-`3`; up to about 125 clicks a second, where 20 ms lost steps above about 12) and timestamps fan tach edges. `loop()` logs expansion-encoder clicks (`EXP1: +1`), which have no function yet, and runs the button logic on the latest sample every 20 ms, so button timing is unchanged. `~INT` stays unconnected: the module's header has no free GPIO, and polling at 2 ms gives the same result for hand-turned knobs.
- **2026-11-10 Mainboard CPL: connectors placed by their centre, `Q5` turned (CPL only; checked against JLCPCB's footprints, not yet in a new JLCPCB preview).** JLCPCB places parts by the origin of its own footprint, the pad centre, but KiCad's connector footprints put the origin on pin 1, so every header was drawn half its length off in JLCPCB's placement preview. `tools/check_cpl_rotation.py` now also works out the move from KiCad's origin to JLCPCB's for each part (`jlc_export/cpl_position_offsets.json`: the 18 connectors and a 0.02 mm nudge on `Q6`) and `tools/build_jlc_assembly.py` adds it to the CPL. `Q5` gets its CPL turn by hand (+270, CPL 0), since its pad numbers differ from JLCPCB's footprint.
- **2026-11-10 One menu on every page; documents tidied (firmware and docs).** Every page the unit serves (live control, simulator, settings, log, firmware update, Wi-Fi setup and the two documents) has the same menu bar: Live, Simulator and Settings always, and a Menu with every page and every settings tab, the current one marked (`src/web_nav.h`; `docs/build_docs.py` puts the same menu on the on-device documents). The setup network's Wi-Fi page keeps a plain title. The documents lost the dated status box at the top of the User Guide and most "was X until <date>" notes (that history is here), and the Hardware Overview's part counts and ring cost are current.
- **2026-11-10 Board respin: 3.3 V from the module, TCA9555, JST VH ring power, ring thermistor, 1 oz (both boards and firmware; ERC/DRC 0, host tests pass, ESP32 compile checked; not tested on hardware).**
  - **3.3 V:** the mainboard's `U2` (K7803-2000R3, 12 V to 3.3 V) and `C12`/`C13`/`C14` are gone. The ESP32-S3-ETH module's own JW5060 buck makes `+3V3` from `VSYS`, which the board now feeds from `+5V_LED_LOGIC` (`D33`); `3V3_EN` is left open. About $2.40-3.40 less per board, the lowest-stock part gone, no 17.5 mm tall part. The logic now needs the 5 V supply, and the module's USB-C powers the logic on its own (the 9-12 V adapter on `J-PWR-12V` no longer does).
  - **Expander:** `U1` MCP23017 (SOIC-28W) -> TCA9555PWR (TSSOP-24, C465732), about $0.65-1.30 less and in deep stock; same address and board pins (port 0/1 = old A/B), re-routed. The firmware detects either chip, so older boards and the breadboard's DIP MCP23017 still work.
  - **Ring power connector:** Mini-Fit Jr 2x3 -> JST VH 4-pin: B4P-VH (C160317) on the mainboard, side-entry B4PS-VH (C157996) on the ring, 1-2 +5 V, 3-4 GND, VHR-4N cable. The ring's `J1` sits in the bottom-right corner with its opening toward the left.
  - **Ring thermistor:** `TH1`, a Murata NCP18XH103F03RB (10k 1 %, B 3434 K, C13564) on the ring just inside the inner LEDs; the signal connectors are now JST XH 5-pin (B5B-XH-A, C157991) with pin 5 = NTC. On the mainboard: `D34` ESD clamp, `R65` 10k 1 % pull-up, `R66` 1k and `C16` 100 nF into GPIO3. `RING_NTC_PIN 3` is now defined, so the ring-temperature firmware is live.
  - **Pin moves:** the main encoder button moved from GPIO40 to the expander (`P0_3`), and the hue encoder's B line from GPIO3 to GPIO40, so GPIO3 (ADC1) reads the thermistor.
  - **Copper:** the mainboard is ordered at 1 oz like the ring (IPC-2221: the 6.5 mm neck carries about 9.2 A at a 10 °C rise, above the 7.5 A fuse).
  - BOM/CPL, Gerbers, renders, schematic PDFs, reports and the breadboard guide regenerated. Check `U1` and the ring `J1` in JLCPCB's placement preview (their CPL rotations could not be looked up on the day).
- **2026-11-10 Ring temperature: fan first, then current (firmware and docs; host tests pass, ESP32 compile checked; needs the board respin with the sensor).** A thermistor on the LED ring (on the signal cable's new pin 5, read on GPIO3) drives the fan up first (45 C to full at 60 C), and only then lowers the ring's current limit (to 1500 mA at 75 C), gradually. An open sensor gives full fan and a 3000 mA limit. Settings on the LEDs tab; temperature and limit on the Status tab, `/status` and the log. Inactive until `RING_NTC_PIN` is defined for the new board.
- **2026-11-10 White background for effects (firmware and docs; host tests pass, ESP32 compile checked).** A new preset setting, `bg=white` (a **Background** box in the preset editor, shown for sweep, segments, spiral, pulse and noise), fills the parts of the pattern that would be off with white instead, at the colour temperature and VAL. Soft edges blend into the white; a palette's own black stays black; the current limits apply.
- **2026-11-10 Preset editor shows only what the effect uses (firmware and docs).** Settings an effect ignores are hidden (solid: none of intensity, count, width, drift; pulse: intensity only; palette and sweep: no count or width; wipe and segments: no width), the colour pickers show only with the *colours* palette, and intensity and count are named for the effect (Twist, Arms, Tail length, Detail, Symmetry and so on). Hidden width and drift are left out of the saved line.
- **2026-11-10 Lit width and drift (firmware and docs; host tests pass, ESP32 compile checked).** Two new preset settings, both in the editor: `width` (0-255, default 128 = unchanged look) sets how much is lit against dark for spiral and noise, the effects whose intensity means something else; `drift` (turns per unit, default 0) moves the pattern's zero point, so a wipe's starting place, the palette's position under a sweep and a palette's seam travel round the ring. Preset lines may now be up to 219 characters and the list 16,000. The editor's Preview checkbox moved to the top, away from Save.
- **2026-11-10 Editor preview no longer flips back to the selected preset (firmware).** The preview was sent as its own request while the page streamed frames on another connection; the web server serves one connection at a time, so the preview could wait for seconds, time out and come back. It now rides on the frame requests (`/sim/frame?pv=1&line=...`), each frame extends it, and Load and Save pause the frame stream while they run.
- **2026-11-09 Preset editor on the simulator page (firmware; host tests pass, ESP32 compile checked).** Under the picture on `/sim`: number, name, effect, palette, speed, unit, intensity, count and three colours, with the finished line shown. **Preview** has the engine render it live, on the ring too (`ledPreviewLine()`, held 5 s per send, so a closed page lets go); **Save** replaces that preset in the running table with no restart (`ledPresetSaveLine()`), and the file is rewritten from the table (`ledpreset::formatAll()`, tested) once the platter stops. Comment lines in the file are not kept by an editor save.
- **2026-11-09 LED simulator page (firmware; ESP32 compile checked).** `/sim` draws the ring live in the browser from the engine's latest frame (`/sim/frame`, 534 bytes, about 12 times a second; LED positions from `/sim/map`), as sent or at full brightness, with or without a ring attached. Linked from the live page; the engine copies about 1 KB a frame for it. `LED_SIM_PAGE` in `pins.h` leaves it out.
- **2026-11-09 Setup network limited to the Wi-Fi page; names on the live page (firmware; host tests pass, ESP32 compile checked).** A request that arrives over the unit's own setup network (local address 192.168.4.1) now reaches only `/wifi`: every other page redirects there and a firmware upload is refused. The setup network's password is the same on every unit and it starts by itself at a venue with no known network, and with no web password set that left the controls and settings open to anyone nearby. The live page shows each preset's and palette's name beside its number (new `GET /names`).
- **2026-11-09 Review fixes and fuller preset and palette docs (firmware and docs; host tests pass, ESP32 compile checked).** The User Guide has a table for the presets (every setting of each), a table and a colour-strip picture for the palettes (`docs/presets/palettes_sheet.png`), and a table of the preset-line keys; two leftovers from the WLED days are gone ("sixteen presets", the preset's own brightness). Built-in preset 5 ("Two arms lava") showed one arm, because the lava palette's first colour is black: it now uses its own three colours. The live page's sliders start at the right limits (19 presets, 10 palettes) before the first status arrives.
- **2026-11-09 Mainboard leftovers from the DevKitC (board; DRC 0 violations, 0 unconnected).** Removed the "USB-C END" silkscreen and its arrow (they marked the DevKitC's orientation), a 0.3 mm GND track stub that ended on nothing, and the GND via it hung from (connected on the bottom layer only). Two other GND stubs that looked loose end inside a connector pad's copper and stay. The power-sheet note no longer mentions the WLED co-processor or the old 5 V fuse reasoning, and two footprint descriptions no longer mention `J6`/the WLED module. Renders, schematic PDF and `Mainboard_gerbers.zip` re-exported (outline identical to the last export). The `+5V_LED_LOGIC` track along the top edge to x = 86 mm is live: it feeds the LED buffers from `D33`.
- **2026-11-09 Smooth brightness (firmware; host tests pass, ESP32 compile checked, bench check pending).** VAL (and the pulse brightness) now follow the CIE lightness curve (`ledfx::lightToLevel()`), so the 100 clicks look even; the engine applies brightness and colour temperature into 16 bits (`ledout::Rgb16`), the limiter runs there, and the SK9822 encoder (`encodeSk9822Hd()`) picks each LED's 5-bit current field per frame so dim LEDs keep about 13 bits of depth instead of 8 (at 1 % the brightest channel has about 80 steps). The current field sets drive current, not PWM, so it still films without flicker. Note: mid-knob is now dimmer than before (half the knob is 18 % of full light, not 50 %); pulse brightness values look the same way.
- **2026-11-09 Preset pictures (docs).** The User Guide shows every built-in LED preset: one sheet of all 19 (embedded) and a two-bar animated WebP of each, rendered by `docs/presets/render_presets.cpp` through the firmware's own `ledfx.h`, `ledpreset.h` and `led_map.h`, so they match what the engine draws. `docs/presets/make_preset_images.py` rebuilds them.
- **2026-11-09 Ring at 1 oz copper, 5 V fused in the harness (docs only; no board change).** A JLCPCB quote put 2 oz on the 203 mm ring at about $3 a board more at 50 to 100 boards (roughly double the bare board). At 1 oz the ring's drop doubles to about 8 mV at 4.5 A and its busiest copper (1.4 A per mm at 9.6 A) stays within 1 oz; `J1` joins the pours solidly. The 5 V side, unfused until now, gets one 7.5 A ATO blade fuse in an inline holder at the supply's +V1 terminal, before the split to `J-PWR-5V` pins 1 and 3 (about $1 to $2, no board assembly). The mainboard stays at 2 oz.
- **2026-11-09 Noise effect (firmware; host tests pass, ESP32 compile checked).** A new LED engine effect, `noise`: soft palette blobs in polar form that turn with the speed and keep changing shape with the beat count; `count` gives rotational symmetry (a kaleidoscope), intensity the detail. Our own gradient noise in `ledfx.h` (`noise3()`), after Animartrix's polar-noise idea without its code. `render()` now takes the beat count rather than the beat phase. Three new built-in presets: 17 Lava noise, 18 Ocean kaleidoscope, 19 Platter noise (locked to the platter).
- **2026-11-09 Product build cleanup (firmware; host tests pass, ESP32 compile checked).**
  - **Bench LED mirror off:** `BENCH_LED_MIRROR` in `src/pins.h` is commented out, so GPIO21 stays the tap button. Uncomment it for a bare-board bench test.
  - **No FastLED:** the eight named palettes are now FastLED's tables copied into `ledfx.h` (`kNamedPalettes`, `namedPalette()`, MIT attribution), tested in `tests/test_ledfx.cpp`; the library is gone from the build and CI.
  - **Open item:** the RD-65A as a smaller supply, to revisit (README, Open items).
- **2026-11-09 Wi-Fi setup without Ethernet; the LED engine page becomes the LEDs tab (firmware and docs; host tests pass, ESP32 compile checked, bench check on hardware).** A new unit could only reach its settings, and so its Wi-Fi, over Ethernet.
  - **Setup access point** (`src/netsetup.cpp`, decisions in `src/setupap.h`, `tests/test_setupap.cpp`): `LiquidTurntable-XXXX`, password `turntable`, at 192.168.4.1 with a captive portal to the new `/wifi` page (network list from a scan, password, Save & Restart). It starts once if no link has an address a minute after boot, or at once after the TAP hold at power-up (which already cleared the password and static IP), and stops when a link has been up for 2 minutes with nobody joined (15 minutes when forced). Station attempts wait while a phone is joined.
  - **Serial command:** `wifi <name> [password]`, `wifi` (status), `wifi -` (forget) on the USB serial port.
  - **No WLED:** the WLED link is gone from the firmware (`wled.cpp`, `wled.h`, `wledmode.h`, the serial/HTTP sender, the WLED target, JSON-serial and highest-preset/palette settings, and the WLED-or-engine choice): the LED engine always drives the ring. Preset mode and the beat flash live in `lighting.cpp`/`presetmode.h` (`tests/test_presetmode.cpp`); `wledPulseBri`/`wledPulseDiv`/`wledPreset`/`wledPalette` are now `pulseBri`/`pulseDiv`/`lightPreset`/`lightPalette`. OSC `/presetmode` replaces `/wledmode`, which still works; `/status` reports `lightMode`.
  - **LEDs tab:** the LED engine page is now a tab of the settings page, saved with every other tab by the one **Save & Reboot** (`ledFormCheck()`/`ledFormSave()`; an unreadable preset line stops the whole save); `/leds` opens it.
  - **Breadboard Guide and README_DETAILED:** the lighting stage now has the ESP32-S3-ETH drive an APA102/SK9822 strip through two gates of a 74AHCT125 (data and clock) from a 5 V adapter, checked against the new schematic (pin map re-exported); `README_DETAILED.md` describes the single-chip design throughout.
- **2026-11-09 Single chip and an SK9822 ring become the standard design (both boards and firmware defaults; ERC/DRC clean, host tests pass, ESP32 compile checked, not tested on hardware).** The dual-chip design (WLED DevKitC co-processor, SKC6812RV ring) is kept on an archive branch.
  - **Ring:** `D1`-`D178` are OPSCO SK9822-A (C5378730; APA102-type, data and clock, 27 kHz PWM, so the ring films without flicker) on a project footprint from the maker's land pattern (`LEDRing.pretty`), with KiCad's stock `LED:APA102` symbol. Same positions for every LED and capacitor; in a row, data and clock are two straight top-layer traces; the 17 row links and the jump under the centre hole are pairs on the bottom layer. New `J2` (JST XH 4-pin, C144395: DATA, GND, CLK, GND) feeds `R1` and the new `R2` (470R each, as the datasheet requires); `J1` (Mini-Fit) is power only. Every LED needs +180 in the CPL, checked against JLCPCB's own footprint. The SK9822-A's supply limit is 5.3 V: keep the 5 V supply at 5.1 V.
  - **Mainboard:** the DevKitC sockets `J6-J1`/`J6-J3` are gone. GPIO18 (data) and GPIO2 (clock), the old WLED serial pins, go through `R18`/`R61` (330R) to `U3`/`U5` (74AHCT1G125) and `R63`/`R64` (33R) to `J-LED-SIG` (JST XH 4-pin), which mates the ring's `J2` in its own cable. `R26`/`R62` hold both low at boot. `D33` now feeds the two buffers (`+5V_LED_LOGIC`). `J5` carries only power. New test point `TP13` (clock); `TP8` and `TP12` are now DAT 3V3 and DAT. The antenna keep-out checker skips the DevKitC when it is absent.
  - **Firmware:** the LED engine is the default (`/leds`): SK9822, data GPIO18, clock GPIO2, RBG, current field 31, 2 MHz. A new **per-LED limit** (`ledout::capLeds`, default 36.6 mA, the SK9822 maker's field-22 heat level) scales any LED that would draw more, before the 4.5 A ring limit. Single colours (18 mA) keep full output, about 47 % brighter than field 22, and white stays where the maker puts it. "mA per LED" is now given at field 31 (54) and scaled to the field. Host tests cover it. The WLED link stays selectable for an older dual-chip board.
  - **Power review:** the ring copper was simulated from the filled pours: about 4 mV to the worst LED at 4.5 A and 8 mV at 9.6 A, the same as the SKC6812RV board. Most of the drop is in contacts and wire. The Hardware Overview gives the wire sizes (power harness 18 AWG or heavier, signal cable 22-24 AWG, whose grounds carry about 1 A of the ring's return) and a brightness estimate for the 1/4 in acrylic platter.
  - **Docs:** README, Hardware Overview, User Guide (the WLED setup appendix is gone; lighting now describes the engine) and the power and signal diagrams. `README_DETAILED.md` carries a note that its WLED sections describe the archived design; the Breadboard Guide still shows the dual-chip lighting stage.
- **2026-11-08 In-firmware LED engine, alongside WLED (firmware; host tests pass, ESP32 compile checked, not tested on hardware).** The ring can be driven from the main chip instead of the WLED DevKitC, selected on the new `/leds` page (default: WLED, unchanged). WS281x/SKC6812RV output on GPIO18 through the RMT peripheral with DMA, or SK9822 (data + clock) over SPI3 with DMA; a per-LED current limiter (36 mA, 4.5 A cap); colour mode, master brightness, colour temperature and the beat flash done locally; seven effects (solid, palette fade/cycle, wipe, sweep, segments, spiral, radial pulse) with speed in turns per bar, per beat or locked to the platter; 16 built-in presets in `/ledpresets.txt` on the FAT partition, picked with the same HUE/SAT knobs, phone page and OSC `/preset` `/palette`, editable on `/leds`; optional Art-Net/E1.31 input of two universes that takes over while frames arrive. FastLED 3.10.3 for palettes only. `tools/gen_led_map.py` generates `src/led_map.h` from the ring board. New tests: `test_ledout`, `test_ledfx`, `test_ledpreset`, `test_artnet`. The ring's data line must come from GPIO18 for this, which the current mainboard doesn't do. Bench checks in the README's open items (STEP jitter with the ring animating, SK9822 colour order).
- **2026-11-08 LED ring current figures and WLED's limit corrected (docs only; not tested on hardware).** Both LED datasheets (SK6812-012 and SKC6812RV) give 12 mA per colour, so full white is about 6.5 A (7.3 A at the maximum drive current), not 10.7 A (60 mA per LED), and they limit three-colour use to 70% of full scale. WLED's setup is now *Max. current for a single LED* Custom 36 mA and *Maximum Current* 4600 mA, which holds the ring to about 4.5 A. The old advice (5500 mA with WLED's 55 mA-per-LED default) held it to about 3.6 A, so the ring can now run about a quarter brighter. Ring-current, connector-pin, `Q5` loss and supply figures follow; the RD-85A's 8 A 5 V rating now covers full white. The board copper notes still quote 10.7-14 A as their design margin.
- **2026-11-08 LED ring: SKC6812RV LEDs, and the LED rotation in the CPL fixed (BOM, CPL and values; not tested on hardware).** `D1`-`D178` move from OPSCO's SK6812-012 (C5378720, 1 kHz PWM, 22,963 in stock) to OPSCO's SKC6812RV (C2890040, 4 kHz PWM, 83,968 in stock), so the ring films with a quarter of the PWM flicker. Both datasheets agree on pinout, 12 mA per colour, GRB order and data timing; the footprint is unchanged. While checking it: JLCPCB's footprint for these LEDs has pad 1 at the opposite corner to KiCad's, and the ring's CPL used KiCad's rotation as is, which would have placed every LED reversed. The ring now has `tools/check_cpl_rotation.py` and its CPL adds +180 to every LED. ERC 0, DRC with schematic parity 0/0/0.
- **2026-11-08 LED ring: every part on the top side (board only; not tested on hardware).** The 178 decoupling capacitors moved from the bottom layer to the top, each 4.1 mm below its LED beside the +5 V pad, with its GND pad on one via to the bottom ground plane; the 178 +5 V stitch vias that fed the old bottom-side capacitors are gone (388 vias either way). One-sided assembly saves JLCPCB's double-sided setup and stencil premium, $33.77 per order. DRC with schematic parity 0/0/0, ERC 0; BOM/CPL, gerbers, PDF and reports regenerated.
- **2026-11-08 Mainboard stock swaps: `D5`, `C6` and the module sockets (schematic, board and BOM; not tested on hardware).** From a cost and stock review against JLCPCB's live library. `D5` moves to MDD's SMBJ15A (C113988, same 15 V standoff and 24.4 V clamp; the GOODWORK part was down to 4,072). `C6` becomes the SMD 100 µF/35 V Lelon already used for `C8`/`C11` (C176683), replacing the through-hole 220 µF Rubycon: one fewer part type and one fewer wave-soldered part; its GND pad sits where the old GND hole was, with two 0.8/0.4 mm vias to the ground plane, and the ring keeps its own 1000 µF `CP1`. The DevKitC rows `J6-J1`/`J6-J3` are now one 1x22 socket each (Hong Cheng C41376161) instead of a 1x20 plus a 1x02, so `J6-J1B`/`J6-J3B` are gone (same pads, same nets; in the schematic `D33` and the pin 21-22 GND symbols moved up to the 1x22's pins). `J1-H1`/`J1-H2` move to Kinghelm C2905423 and `J2-MODE`/`J2-MOTOR` to JXTCONN C42431840, same footprints and 8.5 mm height. ERC 0; DRC with schematic parity 0/0/0; BOM/CPL, CPL rotations, gerbers, reports, PDF and renders regenerated.
- **2026-11-06 Bench LED mirror of the beat flash (firmware; bench only, ESP32 compile checked, not tested on hardware).** `BENCH_LED_MIRROR` (`pins.h`) drives the board's onboard RGB LED on GPIO21 with each WLED flash, at the flash brightness, and hands GPIO21 back to the tap button (`INPUT_PULLUP`) when the flash ends. GPIO21 is also `tapBtnPin`, so a tap during a flash can be missed. Set the flag to 0 (or delete the block in `wled.cpp`) for hardware.
- **2026-11-06 Beat flash follows Pulse's half lock (firmware; host tests pass, ESP32 compile checked, not tested on hardware).** Pulse's half and double locks are toggles, and in a capture, half halved `/bpm` (124 to 62) while the beat stream kept its rate. The platter already follows `/bpm`, so it needed no change. The flash didn't follow, so it now compares the displayed tempo (`BPM`) with the beat rate the sender is sending. The beat clock keeps a smoothed interval between `/beat` events (gaps of 150-3000 ms only; `/measure` doesn't count), and `beat::flashDivider()` doubles the `/pulsediv` divider when the displayed tempo is under 75% of the beat rate. So under half, the flash skips every other beat. Double isn't handled: Pulse doesn't always double the beat stream, so when the rates match, the divider stays as set. Tested in `tests/test_beat.cpp`. Docs: README, README_DETAILED and the User Guide OSC tables.
- **2026-11-06 Beat flash brightness: `/pulse`, the live page and the Show tab (firmware; host tests pass, compiles not checked, not tested on hardware).** The flash on each beat was fixed at full brightness. It now goes to `wledPulseBri` (WLED 0-255), which can be set three ways: OSC `/pulse` (0.0-1.0, like `/value`), the live page's "Beat flash brightness" slider (`/control?a=pulse`), and the Show tab's "Beat flash" field (0-255), which is saved and applies after the reboot the form triggers. `startPulseBri` is the saved default, loaded at boot into `wledPulseBri`, the same pattern as the colour temperature. `/status` reports `pulseBri`. `wledmode::pulseFromUnit()` converts and clamps 0.0-1.0 to 0-255; tested in `tests/test_wledmode.cpp`. Setting 0 turns the flash off. Docs: README and README_DETAILED OSC tables, the User Guide OSC table, and the rebuilt HTML and copies.
- **2026-11-06 Beat flash divider: `/pulsediv` (firmware; host tests pass, compiles not checked, not tested on hardware).** The flash fired on every beat, which is 120 a minute at 120 BPM. `beat::flashOnBeat()` now flashes on beats 0, N, 2N… of the bar, using the `/beat` index, so 4 gives the downbeat only in 4/4. The divider is `wledPulseDiv` (1-16, default 1 = every beat). OSC `/pulsediv` sets it live; the Show tab's "Beat flash divider" field (`startPulseDiv`, prefs `pulsediv`) sets the value kept across reboots, the same pattern as the brightness. `/status` reports `pulseDiv`. A bad divider falls back to every beat. Tested in `tests/test_beat.cpp`. Docs: README, README_DETAILED and the User Guide OSC tables.
- **2026-11-06 Beat keep-alive, and `/measure` as bar position (firmware; host tests pass, compiles not checked, not tested on hardware).** Pulse's feed, checked on the wire at 120 BPM: `/beat` is an int, the beat index in the bar (0 on the downbeat), about 2 a second; `/measure` is a float, the bar position in beats, about 30 a second, wrapping to 0 on each downbeat; `/bpm` is sent only on change; `/confidence` is an int, 0 throughout. The firmware had treated every `/measure` as a downbeat, which reset the beat count about 30 times a second. Now `/beat` sets the index from its argument (no argument still counts up), and `/measure` sets the phase only (`beat::positionEvent()`, anchored to the position's fraction). Keep-alive: `/bpm` times out after 60 s, and Pulse sends it only on change, so a steady tempo would fall back to the tap. A beat or `/measure` from the chosen sender now refreshes that source's freshness stamp (`beatKeepAlive()` in `osc.cpp`). It doesn't change the BPM value, and the usual timeouts apply once the beats stop. Docs: README and README_DETAILED failover and OSC rows, the User Guide OSC table and stale-feed note.
- **2026-11-06 Beat gate follows the failover, not the sender's IP (firmware; host tests pass, compiles not checked, not tested on hardware).** Replaces the IP-owner gate below, which couldn't separate two senders on one machine. `setBpmSource()` (`tempo.cpp`) is now the only place `bpmSource` changes. The failover code (the `/bpm` and tempo-controller takeovers, the tap, and `loop()`'s timeouts) calls it. It sets `beatSenderChosen`, which opens the beat gate for an OSC source and closes it for the tap, and it clears the beat clock whenever the source changes. `/beat` and `/measure` (and the WLED flash) are gated on that flag. `beat::fromOwner()` and the owner state are removed. Caveat: the gate doesn't identify the sender, so two live senders both count. Also, `loop()` chooses the tempo controller when `/bpm` times out without checking the tempo feed is live, so a beat in that window still counts.
- **2026-11-06 Beats follow the BPM owner (firmware; host tests pass, compiles not checked, not tested on hardware).** `/beat` and `/measure` now use the same failover as BPM. The sender that set the tempo is the owner: the last `/bpm`, or the tempo-controller message that took over. `setBpmOwner()` (`tempo.cpp`) records its source and IP, and clears the beat clock whenever the owner changes. Beats from any other sender are ignored, including their WLED flash, so Resolume's beats can't move the clock or flash the ring while `/bpm` is live. The owner is matched by IP only, so two senders on one machine count as one. `beat::fromOwner()` in `beat.h` is tested in `tests/test_beat.cpp`. Docs: README, README_DETAILED and the User Guide OSC tables and failover section; the documents and their HTML are rebuilt, and the stale `docs/files/` copies (enclosure, mainboard JLC notes) are re-synced.
- **2026-11-06 Beat clock from OSC `/beat` and `/measure` (firmware; host tests pass, compiles not checked, not tested on hardware).** New `src/beat.h` (pure, no Arduino headers) with `tests/test_beat.cpp`: a count of beats since the downbeat, the phase within the beat (from the BPM in use), and a freshness check (the last beat is within two periods). `/measure` is the downbeat and resets the count. A `/beat` within 30 ms of a `/measure` is the same downbeat, so Resolume can send either first. Both are ignored while the tap is the tempo source, so the clock follows whichever OSC sender drives BPM: the show software or Resolume's tempo feed. If both send beats, the last one wins. `/status` gains `beatFresh`, `beatPhase` and `beatInMeasure`. The clock is not used by the motor yet.
- **2026-11-06 WLED beat flash (firmware; host tests pass, compiles not checked, not tested on hardware).** Each OSC `/beat` message flashes the WLED ring to full brightness for 80 ms, then sends the master brightness again. The message is `{"bri":N,"transition":0}`, so WLED jumps rather than fades. Serial path only: an HTTP send can block `loop()` for about 0.6 s and would land the flash a beat late, so on a networked WLED the flash is dropped. `onBeat()` (core 0) only sets a flag; `wledService()` (core 1) sends the flash, so no serial write happens off core 1. The payload builder and `flashHoldMs` are in `src/wledmode.h`, tested in `tests/test_wledmode.cpp`. `/measure` is not used yet. To check on a bench: a flash per `/beat` at 120 BPM, and no fade or stutter with a colour change during a flash.
- **2026-10-25 review fixes: the 5 V side stays unfused, documents corrected, `prefs` race fixed (docs and firmware; compiles, not tested on hardware).** A review found that the reason given for leaving the 5 V side unfused (the RD-85A limits its own 5 V current below 15 A) is not in the supply's datasheet: its overload protection is 110 to 150% of total rated power (88 W), not a 5 V current limit. The 5 V side stays unfused by choice, and the documents now say so. Corrected: the power sections of `README.md` and `README_DETAILED.md` (F3, the removed crowbar `Q7`/`D34`/`R60` and its bench step, `+5V_LED_RINGFUSE`, the claim that each rail has a fuse); the panel-line count (21 = CLK and DT of 7 encoders, SW of 6, plus TAP_BTN); the smooth-motion floor text (default 2.0, a feel setting, as in `state.h`); the heat section's box dimensions (the figures were made for the earlier 14 x 17 x 2 in box and have not been redone); the Hardware Overview's part and joint counts (SMD 77 parts and 235 joints, full 100 and 397, from `build_jlc_assembly.py`); the ring-board order notes (a current-state block at the top, and the 10-17 and 10-18 entries marked superseded); the breadboard guide's 2N7000 (TO-92) is labelled a stand-in for the board's SOT-23 2N7002; the `layout.scad` outside-size comment. Firmware: `prefs` is one NVS object used from core 0 (web saves) and core 1 (deferred saves in `loop()`) with no lock; every begin/end pair now takes `prefsMutex`, created first in `setup()`. CI: a step regenerates the breadboard guide from its committed pin map and fails on any difference. The pin map itself still needs `kicad-cli`, so it stays a local check.
- **2026-10-21 Web UI screenshots for preset mode (docs and a one-line firmware change).** New `docs/screenshots/webui_live.png` and re-made `webui_status.png` and `webui_show.png`, shown in the User Guide and the index. They are rendered in Chrome from the firmware's own page code (`handleConfigRoot()` and the `/live` page, run on a PC with sample values, a unit in preset mode with preset 7 and palette 12), not captured from a unit, so the build date on the Status page is the day they were made. The Status tab now has a *Lighting* line. Other screenshots (Network, Colour, Motor, Security, log, firmware update) are unchanged.
- **2026-10-20 WLED preset mode (firmware; compiles and the PC tests pass, not tested on hardware).** Holding the HUE button for 1 s switches between colour mode and preset mode. In preset mode the HUE knob selects a WLED preset (1 to the configured top), the SAT knob a palette (0 to the configured top), both wrapping, and VAL stays the master brightness; a short HUE press returns to preset 1 and a SAT press to palette 0. The unit sends `{"ps":N}`, then 300 ms later `{"bri":B,"seg":[{"pal":L}]}` (a preset loads its own brightness and palette); leaving the mode sends one colour message that also selects Solid and palette 0, as chosen. A short HUE press now acts on release (it has to wait to tell a press from a hold). The tops are two new settings on the Show tab (`wpresetmax` 1-250, default 16; `wpalmax` 0-255, default 70) because WLED cannot report them over serial. The mode is runtime only: the unit starts in colour mode. Also in `/status` (`wledMode`, `preset`, `palette`, `presetMax`, `paletteMax`), `/control` (`mode`, `preset`, `palette`), OSC (`/wledmode`, `/preset`, `/palette`) and the phone page. New `src/wledmode.h` (selector wrap and the JSON builders) with `tests/test_wledmode.cpp`; `wled.cpp` reworked around one `wledSend()`. Docs: User Guide *Preset mode* section, panel, phone page, Show tab, OSC and troubleshooting entries; both READMEs; Hardware Overview. The Status tab gained a *Lighting* line (colour mode, or preset mode with preset and palette). To check on a bench: that WLED re-applies the same preset id after the Solid message, and that the palette knob overrides a preset's own palette.
- **2026-10-20 Documentation audit against the current design (docs only).** `power-flow.svg` now also shows the `Q5` reverse-polarity FET, the `Q7` crowbar and the `D33` feed to the WLED DevKitC (which has its own regulator), and the 3.3 V loads no longer list the DevKitC; `signal-flow.svg` now shows the fan (PWM from the main controller, tach into the expander) and what the expander carries. README, README_DETAILED, Hardware Overview and User Guide: the new Test points sheet and a Test points section; six function sheets, not five; the Hardware Overview's sheet table and key parts list `Q5`, `Q7`, `D32`, `D33` and `Q6`, its supply description counts the DevKitC as a 5 V load, its ordering counts are the current 32 types / 83 parts / 258 joints (SMD) and 106 parts / 420 joints (full), and the known-limits line no longer puts the WLED chip on the 3.3 V rail; the User Guide now names both power connectors, no longer warns that USB into the lighting controller back-feeds the 3.3 V supply (it has not since 2026-10-17), and gives the WLED current limit as 5,500 mA as the README does; the flash size of the embedded documents is about 200 KB; README_DETAILED's repo layout lists the docs index, `docs/files/`, the breadboard guide, `tinkercad/` and the new tools; `ASSEMBLY_COMPARE.md` counts updated. History entries and the dated `jlc_export/README.md` log are left as written.
- **2026-10-20 `power-flow.svg` redrawn for the split power input (docs only).** The diagram still showed one 8-pin `J-PWR-IN`; it now shows `J-PWR-12V` (JST XH, 2 pin) on the 12 V side and `J-PWR-5V` (Mini-Fit Jr, 4 pin) on the 5 V side, as split on 2026-10-17, and `F1` reads 2 A (it said 2.5 A). The diagrams are also published as copies in `docs/files/diagrams/`, which the documentation index links to.
- **2026-10-20 Folder index pages in docs/ (docs only).** `docs/sync_files.py` now generates a dark-themed `index.html` in every folder under `docs/` that had none (`breadboard`, `diagrams`, `screenshots`, and the folders under `files/`), listing subfolders, files with type and size, and thumbnails for images, so a folder URL no longer answers Forbidden. `--check` fails if one is stale or missing.
- **2026-10-20 Twelve test points on the main board.** Three GND loops and pads on `+12V_MOTOR`, `+12V_MOTOR_F1`, `+3V3`, `+5V_LED_RING`, `WLED_LINK_TX`, `I2C_SCL`, `I2C_SDA`, `ENABLE` and `LED_DATA`, silkscreened with the net, in a new `testpoints.kicad_sch` sheet (`TP1` to `TP12`, global labels, not in the BOM or placement files) and placed by `kicad_mainboard/tools/add_test_points.py`. Each pad sits on copper of its own net, so no tracks changed. They are kept out from under the plugged-in modules; `STEP`, `DIR` and the TMC2209 UART lines run only beneath them and have no pad. ERC 0; DRC 0 violations, 0 unconnected (the zones were refilled, so the diff of `Mainboard.kicad_pcb` is mostly fill polygons). Gerbers and drill files re-exported (same 12-file set), schematic PDF, renders, ERC and DRC reports regenerated; DRC with schematic parity 0 violations, 0 parity issues. The BOM and CPL are unchanged (the test points are excluded from both). Not tested on a built board.
- **2026-10-18 `U1` moved from the MCP23017-E/SP (DIP-28) to the MCP23017-E/SO (SOIC-28W, LCSC C47023).** Saves about $1.05 a board at 100 boards (BOM $8.79 to $7.92; $12.63 to $11.22 for one board) and moves the part from the THT to the SMD assembly pass (2,666 in stock against 259). Same pinout, so no schematic nets or firmware change. The 2.54 mm DIP footprint had tracks running between its pins, so the panel-area copper around it was reworked: the 12 expander signal nets, `I2C_SDA`/`I2C_SCL` and the `+3V3`/`GND` links were ripped up and re-routed from escape vias at each SOIC pad with `tools/microroute.py` (its clearance margin raised from 0.22 to 0.26 mm), `SAT_DT`, `SAT_CLK`, `HUE_DT` and `HUE_CLK` were re-routed around the smaller footprint, `U1` sits 2.3 mm higher, and `C9`'s ground now goes through its own via. ERC 0; DRC with schematic parity 0 violations, 0 unconnected, 0 parity issues; BOM/CPL, gerbers, reports, PDF and renders regenerated. CPL rotations checked against JLCPCB's own EasyEDA footprints (pad pattern matched by pad number, `tools/check_cpl_rotation.py`) and corrected: `jlc_export/cpl_rotation_offsets.json` now holds the turn to add to KiCad's rotation for every part type, and `build_jlc_assembly.py` applies it (`U1` and `U3` 0 degrees, `Q6` and `C8`/`C11` +180, the module and encoder headers +270, `J3`/`J5`/`J-PWR-12V` +180, `U2` +270, passives and diodes unchanged). Not decided, so still KiCad's rotation: `Q5`, `Q7` (different pad numbering) and the four fuse clips; check those in JLCPCB's preview. Run the panel I2C check (all encoder buttons, the fan tach and the expansion headers) on the first board.
- **2026-10-21 LED ring: the pin 1 data trace is re-routed and tapered.** The single 0.2 mm trace from `J1` pin 1 left the pad diagonally toward the cable's side. It now leaves straight up from the pad at 0.5 mm, steps to 0.35 mm and then 0.2 mm over about 12 mm, and rejoins its old end point (`R1`'s side). A straight back-side connector was considered and not adopted (see the Hardware Overview, strain relief at `J1`). Ring ERC 0; DRC with schematic parity 0 violations, 0 unconnected, 0 parity issues; ring gerbers, reports and PCB PDF regenerated (the PCB PDF is exported with F.Cu, B.Cu and both silkscreens, which may differ slightly from the earlier one).
- **2026-10-23 enclosure model: 6 mm Baltic birch, ball-unit platter support, sliding drive (draft, untested; docs and CAD only).** Floor, walls and lid are 6 mm (was 12.7 mm); the interior (334.6 x 256.6 mm) is unchanged, so the outside is 346.6 x 268.6 x 80.9 mm, and the platter rim is now only 0.9 mm from the rear edge. The lid top is at 74.9 mm and H is 68.9 mm, so the platter underside, LED face, motor stack and the 3.6 mm plate-to-tray gap are unchanged; `CP1` now fits standing (17.7 mm of headroom). Three inexpensive flanged ball units (flange 14.5 to 14.7 mm, ball 2 to 2.3 mm above the flange, nylon unit rated 2 kg each) are pressed into 12.7 mm placeholder holes on a 118 mm radius, replacing the removed wheel cartridge; the platter drive is a D-flat stub the hub slides on. New `enclosure/sliding_drive_sketch.svg`, `enclosure_balls.stl`, regenerated STLs and renders, and notes. Hardware Overview says the platter support is now documented.
- **2026-10-23 transport case: a custom ATA-style roadcase instead of a Pelican (docs and drawing only).** `enclosure/case_fit.svg` now shows a roadcase with a 455 x 400 x 200 mm interior (50 mm lid, 150 mm body): 53.0 mm of foam at each end, 53.2 mm each side (even on all four plan sides) and 46.1 mm top and bottom around the 349 x 293.7 x 107.9 mm envelope; estimated exterior about 475 x 420 x 226 mm, empty weight about 7 to 8 kg (estimates from typical construction, not a quote). It replaces the Pelican 1550 fit (interior about 468 x 356 x 194 mm, 30.9 mm each side), which is no longer the plan.
- **2026-10-23 enclosure: platter clearances 15 / 15 / 15 / 80 mm, corner blocks, and a case fit (CAD and docs only).** The box is now sized from the platter rim's clearance to the outside edge: 15 mm on the left, front and rear and 80 mm on the control side, so the outside is 349 x 284 x 80.9 mm (it was 346.6 x 268.6, with only 0.9 mm at the rear). The interior is 337 x 272 mm and every part shifts together, so nothing moves relative to the motor; about 20 mm of free space now sits behind the fan and PSU. Four 10 mm square corner blocks, full interior height, take the lid screws (four 3.5 mm lid holes). Overall envelope with 12 mm feet (assumed), the knobs and the rear connectors: 349 x 293.7 x 107.9 mm. The smallest standard Pelican with at least 30 mm of foam all round is the 1550 (interior about 468 x 356 x 194 mm, from retailer tables that disagree by up to 0.2 in): see `enclosure/case_fit.svg`.
- **2026-10-21 breadboard guide brought in line (docs only).** Its "what the breadboard leaves out" table and intro no longer mention the 15 A blade fuse, the fuse clips or the over-voltage circuit, and name the 2 A resettable fuse and the unfused 5 V side. The wiring is unchanged: `export_pinmap.py` and `build_breadboard_guide.py` re-run, 47 nets and 84 wires verified against the schematic.
- **2026-10-21 overvoltage crowbar removed (`Q7`, `D34`, `R60`).** Datasheet check (MDD BZT52C4V7/C6V8 Rev 2025A6, MSKSEMI AON6312-MS, MDD SMBJ5.0A): the circuit is a shunt clamp, not a crowbar, because `Q7`'s gate voltage comes from the rail it pulls down, so it settles at about the zener voltage plus the 1.2 to 2.5 V gate threshold (6.0 to 7.4 V with the 4.7 V zener, 7.7 to 9.8 V with a 6.8 V one), which is where `D32` (SMBJ5.0A: 6.4 to 7.0 V breakdown, 9.2 V at rated surge) already clamps. With a 5 A-limited 12 V supply it would dissipate about 40 W and likely fail short. The 4.7 V zener's worst-case start (about 5.4 V hot) was also close to the supply's 5.5 V limit. Removed from the schematic and board; the connectors stay keyed and of different families and the harness is still metered before it meets the board. ERC 0; DRC with schematic parity 0 violations, 0 unconnected, 0 parity issues; BOM/CPL, gerbers, reports and docs regenerated. Mainboard BOM $10.51 / $9.46 / $7.47 at 1 / 10 / 100 boards.
- **2026-10-21 `F3` removed: the 5 V side has no fuse.** With the chosen supply the fuse could never blow: the RD-85A is rated 8 A on 5 V (10 A range) and goes into overload hiccup below 15 A, and the ring's worst case is within that (the BOM's SK6812-012 is a 12 mA-per-channel part, so 36 mA a LED, about 6.4 A for 178; the docs keep the conservative 60 mA, 10.7 A). `F3A`/`F3B` are gone, the `+5V_LED_RINGFUSE` net is merged into `+5V_LED_PSU` (the two pours are joined and `Q5`'s drain sits on the 5 V input), and the board needs no blade fuse at all. With a larger supply, fit an inline blade-fuse holder in the 5 V harness (documented). ERC 0; DRC with schematic parity 0 violations, 0 unconnected, 0 parity issues; BOM/CPL, gerbers, reports and docs regenerated. Mainboard BOM $10.91 / $9.86 / $7.80 at 1 / 10 / 100 boards. The T508VS clip stock risk is gone with the part.
- **2026-10-21 `F1` is a 2 A resettable fuse (RUILON SMD1812P200TF16, LCSC C20812), not a blade fuse in two clips.** The T508VS clip (C41371869) has only 250 in stock at LCSC and JLCPCB; with `F1`'s two clips gone, `F3` needs two per board, so 100 boards use 200. `F1A`/`F1B` are removed and `F1` is one 1812 part between the same two nets (2 A hold, 3.5 A trip, 16 V, about $0.09; a PTC trips over seconds and resets once the fault is gone, and its hold current falls when hot, so about 1.5 A at 60 C; use the 2.6 A part C20814 if the motor rail ever needs more). `F3` is unchanged. ERC 0; DRC with schematic parity 0 violations, 0 unconnected, 0 parity issues; BOM/CPL, gerbers, reports and docs regenerated. Mainboard BOM $11.13 / $10.08 / $7.95 at 1 / 10 / 100 boards. Confirm C20812's JLCPCB library type and stock at upload.
- **2026-10-18 fuse clips: order in batches (docs only).** The T508VS clip (C41371869) has 250 in stock, enough for 62 boards. No second source exists on LCSC: the through-hole ATO holders are either low in stock, MINI-size, or (Littelfuse 178.6165.0002, C207061, 1,056 in stock) about $3.50 per board dearer and a layout change. Decision: keep the SMD clips and order in batches of up to 60 boards.
- **2026-10-18 stock swaps: `J2-MODE`/`J2-MOTOR` and `U3`.** `J2` sockets moved from Ckmtw C124416 (0 in stock) to HCTL PM254-1-08-Z-8.5, **C2897371** (10,158 in stock, same 2.54 mm pitch and 8.5 mm height, no layout change). `U3` moved from the Nexperia 74AHCT1G125GW (C12495, SC-70-5, 30 in stock) to the TI SN74AHCT1G125DBVR, **C7484** (SOT-23-5, 26,335 in stock; identical pinout, so the nets are unchanged). The SOT-23-5 footprint is larger: `U3` moved 0.3 mm right and 0.2 mm up, `C10` 0.2 mm down, and the five `U3` tracks were re-attached. ERC 0; DRC with schematic parity 0 violations, 0 unconnected, 0 parity issues; BOM/CPL, gerbers, reports, PDF and renders regenerated. Mainboard BOM $12.63 / $11.38 / $8.79 at 1 / 10 / 100 boards. `U3`'s CPL rotation is corrected to 0 (see the `U1` entry). At 100 boards the only short part is now the fuse clip C41371869 (250 vs 400).
- **2026-10-20 Tinkercad page and ZIP; back links; I/O panel restored (docs and enclosure files).** `docs/files/enclosure/tinkercad/index.html` (generated by `docs/sync_files.py`) lists the Tinkercad files with a ZIP bundle at the top, and the documentation index links to it instead of the bare folder, which answered Forbidden; `--check` now rejects links to folders. The two guides' top bars and the Breadboard Guide have an "All documentation" link back to the index (not in the copy the unit serves). `enclosure/layout.scad` had been left with the I/O plate on the inside face of the rear wall by an edit that was meant to be dropped and was committed in 0dc5746; it is restored to the documented outside-plate version and the STL files, renders and Tinkercad page are regenerated from it.
- **2026-10-20 docs/ is self-contained (docs and tools only).** Every file the documents and the index link to now sits inside `docs/`: `docs/sync_files.py` copies the READMEs, the changelog and the board and enclosure files into `docs/files/` and rewrites relative links in the copied Markdown (links to uncopied repository files become GitHub URLs). `--check` fails if a copy is stale or any link in `docs/` points outside it; CI runs it. The Hardware Overview, the Breadboard Guide and the index now link into `docs/files/`. Also fixed `device_copy()` in `build_docs.py`: the on-unit pages replaced links to files that are not on the unit with a stray control character instead of the link text.
- **2026-10-20 Platter support wheels removed from the docs (docs only).** Roller bearings were chosen instead, so the Hardware Overview no longer describes the six ball transfer units or the four-wheel cartridge, `enclosure/WHEEL_SUPPORT.md` is deleted and the documentation index no longer lists it. The earlier changelog entries about the wheels stay as history. The CAD files (`wheel_cartridge.scad`, `cartridge.stl`, `tire.stl`, `mortise.dxf`) are left in `enclosure/` unreferenced. The 2 mm platter height used in the motor-mount stack is now marked as assumed.
- **2026-10-20 Documentation index (docs only).** `docs/index.html`: one dark-themed page (same palette and layout as the generated guides) that lists every document, HTML guides first, then links to the Markdown, PDF, SVG, CSV, DXF, STL and PNG files for the boards, the order files, the enclosure and the web page screenshots, with a filter box. Written by hand; not part of `build_docs.py`.
- **2026-10-20 Motor mount and enclosure layout drafts (enclosure/, docs only).** `motor_mount.scad` (3 mm aluminium plate on four 40 mm standoffs, `motor_plate.dxf`), `layout.scad` and renders of a reduced 14.2 x 11.1 in layout with a 120 mm fan in the floor, a 2.5 in interior and the LED tray hung from the lid, plus `MOTOR_MOUNT.md` with the height stack and assumptions to check. Tinkercad-ready STL files for every part (`enclosure/tinkercad/`, regenerated and mesh-checked by `export_stl.py`), and the plate as SVG and DXF. Rear-wall I/O panel (2 mm aluminium plate: fused and switched IEC inlet, etherCON jack, USB feed-through) added to the layout; IEC and D-size cutouts are placeholders until the real parts are measured. Untested; sizes read off a photo of the prototype.
- **2026-10-19 Breadboard guide (docs and tools only).** `docs/BREADBOARD_GUIDE.html`: a step-by-step, illustrated guide to prototyping the controller on three breadboards (brain, panel, lighting), with a wire table and a test for each stage. It is generated by `docs/breadboard/build_breadboard_guide.py`, which first rebuilds the breadboard's connectivity from its own hole tables and checks it against the schematic netlist (`docs/breadboard/pinmap.json`, exported by `export_pinmap.py`), so a wrong pin or a short stops the build. The breadboard leaves out the fuses, 3.3 V converter and protection parts (listed in the guide).
- **2026-10-19 Six Extended parts swapped for Basic equivalents.** `C1590`→C14663 (`C7`, `C9`, `C10` and the ring's 178 caps), C602037→C45783 (`C12`, `C14`), `R18`→C23138, `R19`/`R26`/`R60`→C25804, `R21`/`R22`→C23162, `R9`/`R23`→C21190. Same footprints, no board change; BOM/CPL regenerated and mainboard parts repriced. Details in `kicad_mainboard/jlc_export/README.md`.
- **2026-10-17 firmware split into modules (no behaviour change intended; compiles, not tested on hardware).** The 3,100-line sketch is now `setup()`/`loop()` plus `src/` modules (motor, panel, tempo, osc, net, web, ota, config, wled, monitor); shared globals are declared once in `src/state.h`, pins in `src/pins.h`. The four copy-pasted encoder ISRs became one `quadIsr()` that reads the GPIO input register instead of calling `digitalRead()`; the decoder is in `src/quadrature.h` with PC tests (`tests/test_quadrature.cpp`, 34 checks, including agreement with the old decoder on all 16 pin-state pairs). Flash 1,410,322 B (was 1,411,142), RAM 68,616 B (was 68,592). CI's firmware job was missing `colour.h` and `src/` and would not have built; fixed. `git gc` brought `.git` from 63 MB to 38 MB.
- **2026-10-18 cost and stock swaps: `Q5`, `C6`, `D1`.** ERC 0; DRC with schematic parity 0 violations, 0 unconnected, 0 parity issues; BOM/CPL, gerbers, reports, PDF and renders regenerated; not tested on hardware. `Q5`: AON6403 (C2760089) replaced by the Siliup SP30P03NK (C41355006), $0.35 against $0.56 at 100 boards, stock 3,734 against 1,465, no lifecycle warning. Datasheet checked: 30 V, +-20 V gate, Vgs(th) -1 to -2.5 V, pinout S S S G / D x4 + tab (same as the footprint), 4.5 mOhm typ but **6 mOhm max** at -4.5 V (the AON6403's maximum is 4.3 mOhm), so the worst case at the ring's 10.7 A rises from about 0.5 W to about 0.7 W (0.18 W at the 5.5 A WLED limit). Footprint unchanged. `C6`: SamYoung NXH35V220 (C346959, 375 in stock) replaced by Rubycon 35ZLH220MEFC8X11.5 (C107639, 40,925 in stock, $0.08 against $0.10): datasheet checked, same 220 uF/35 V, 8 x 11.5 mm, 3.5 mm lead pitch, 0.6 mm leads, 945 mA ripple and 0.056 ohm at 100 kHz, 8000 h at 105 C, so a straight swap on the same footprint. `D1`: the through-hole 1N5822 (DO-201AD) replaced by an SS34 in SMA (C8678, the same part as `D33`): same 3 A / 40 V, about $0.12 less and one fewer wave-soldered part. New SMA footprint centred at (15.35, 29.0) with 1 mm tracks to the old pad points, so the 12 V routing is unchanged. `U1` (MCP23017 DIP-28 to SOIC-28, C47023, about $0.87 cheaper) was attempted and backed out: the 1.27 mm pads collide with the signals that run between the old 2.54 mm pins, so it needs a proper re-route of the panel area (Freerouting is not installed here).
- **2026-10-18 crowbar threshold corrected (docs and value text only).** The AON6312-MS datasheet gives the gate threshold as 1.2 V minimum and 2.5 V *maximum* (the LCSC listing's 2.5 V is the maximum, not typical), with -4 mV/C. With the 4.7 V zener (+-5%, softer at low current) the crowbar therefore starts somewhere between about 6 V and 7.5 V, not at 7 V, and at the 5.5 V trim limit the gate sits near the minimum threshold, so a hot, low-threshold part may leak a little there. Nothing was changed electrically; a smaller zener was considered and rejected for that reason. Bench-test the start voltage and the 5.5 V leakage.
- **2026-10-17 overvoltage crowbar on the 5 V input: `Q7` (AON6312-MS), `D34` (BZT52C4V7), `R60`.** ERC 0; DRC with schematic parity 0 violations, 0 unconnected, 0 parity issues; gerbers, BOM/CPL, reports, schematic PDF and renders regenerated; **not tested on hardware.** `Q7`, a 30 V N-channel FET (MSKSEMI AON6312-MS, C17702951, 3.2 mOhm at 4.5 V), sits across `+5V_LED_PSU` (drain on the rail, source on GND through three vias to the B.Cu plane) in the pour right of `J-PWR-5V`, at (27.4, 8.0). Its gate is fed from the rail through `D34`, a 4.7 V zener (C173408), with `R60` (10 k) from gate to GND: it stays off at 5.5 V (the gate sits near 0.8 V), starts conducting at about 7 V and is hard on by about 10 V, so a higher voltage on the 5 V pins is pulled down to a few tenths of a volt and the supply trips into its overload hiccup. Same land pattern as `Q5` (`DFN-8_5x6_Single_GSD`). Why a crowbar and not a second back-to-back FET: a FET switched off still passes current through its body diode in one direction, so cutting off overvoltage needs a second FET in series; that doubles the loss (about 1 W total at the ring's 10.7 A, the heat `Q5` was removed for in October) and needs the congested pour corridor re-laid. The crowbar has no series loss and protects the whole `+5V_LED_PSU` net, including the DevKitC feed. Limits, also in the Hardware Overview: a few microseconds of overshoot before `Q7` turns on, no protection between about 5.5 V and 7 V, and it relies on the supply limiting or tripping (RD-85A: overload hiccup); a stiff 12 V source would burn `Q7`. `D33`, `Q7`, `D34` and `R60` have hidden reference text on the board to avoid silkscreen clashes; `D33` moved to (33.4, 8.0) and `J-PWR-5V`'s reference text to (26.5, 16.0). Mainboard component cost per board: $13.24 / $11.91 / $9.15 at 1 / 10 / 100 boards.
- **2026-10-17 power input split into `J-PWR-5V` and `J-PWR-12V`; WLED DevKitC fed through `D33`.** ERC 0; DRC with schematic parity 0 violations, 0 unconnected, 0 parity issues; gerbers, BOM/CPL, reports, schematic PDF and renders regenerated; **not tested on hardware.** **Connector split:** the 8-position Mini-Fit `J-PWR-IN` is replaced by `J-PWR-5V` (Mini-Fit Jr 2x2, HCTL HC-5566-2*2A-05, C20539147; pins 1 and 3 +5V, 2 and 4 GND, same origin so the 5 V pour connections stay) and `J-PWR-12V` (JST XH 2-pin B2B-XH-A, C158012; pin 1 +12V, pin 2 GND, at (8.8, 14.0) rotated 90 degrees, joined to the existing F.Cu run to `F1A` by a 1 mm track). They are different connector families on purpose, so a 12 V plug cannot go into the 5 V header; a wire in the wrong cavity of the 5 V plug is still possible, so the harness test stays. The 12 V B.Cu track and via from the old pin 5, the five thin `+5V_LED_PSU` links to the old pin 3, and the four silkscreen labels of the removed pins were deleted; `J-PWR-5V`'s reference moved below the connector and `J-PWR-12V`'s is hidden. Both connectors sit ahead of the fuses, as before. **DevKitC feed:** `D33` (SS34, C8678, SMA) at (32, 8) takes `+5V_LED_PSU` (its anode sits in the 5 V pour) and its cathode (new net `Net-(D33-K)`) runs on a 0.5 mm F.Cu lane at y = 2.4 above the socket tops, down at x = 86 to `J6-J1B` pin 1 (the DevKitC's 5V0). `J6-J1` pins 1 and 2 (3V3) are now open (no-connect flags, the three F.Cu tracks of their branch removed), so the DevKitC runs from its own regulator: USB into the DevKitC can no longer back-feed `+3V3`, and the 3.3 V rail loses the DevKitC's load. Why `+5V_LED_PSU` and not the ring side: it is the only 5 V copper near the top lane (the ring side is at the far bottom left). It is ahead of `F3` and `Q5`, so the feed is protected by `D33` against a reversed supply but only by the supply's own limit against a short; the 0.5 mm lane carries about 0.3 A. Not checked: how the DevKitC-1 joins its USB 5V to its 5V0 pin, and the dropout margin (about 4.7 V in, its regulator needs about 4.4 V at 0.3 A). Mainboard component cost per board: $12.83 / $11.50 / $8.82 at 1 / 10 / 100 boards.
- **2026-10-17 5 V reverse-polarity protection restored: `Q5` (AON6403) and `R19`.** ERC 0; DRC with schematic parity 0 violations, 0 unconnected, 0 parity issues; gerbers, BOM/CPL, reports, schematic PDF and renders regenerated; **not tested on hardware.** A P-channel FET in the ring's 5 V line: drain on the fuse side (new net `+5V_LED_RINGFUSE`, from `F3B`), source on the ring (`+5V_LED_RING`), gate to GND through `R19` (10 k). Reversed input leaves the body diode and channel off. Part: AOS AON6403 (LCSC C2760089, 30 V, 3.1 mΩ at -10 V, 4.3 mΩ max at -4.5 V, +-20 V gate), about 0.5 W at 10.7 A against about 0.8 to 1.1 W for the old SQJ407EP; 1,465 in stock, one third-party page lists it not recommended for new designs (second source: VBsemi AON6403-VB, C7524994). Placed in the pour corridor below `F3B` at (9.8, 57.0), rotated 90 degrees; the old single ring pour is split into a fuse-side pour (to y = 58.4) and a ring-side pour (from y = 58.9), 0.5 mm apart, and the pour bump east of the old neck was dropped. New footprint `lib.pretty/DFN-8_5x6_Single_GSD` (KiCad's PowerPAK SO-8 Single land pattern, pads renumbered 1 = G, 2 = S, 3 = D to match `Q_PMOS_GSD`); its land pattern is not verified against AOS's drawing (the datasheet in the repo has none), and the JLCPCB CPL rotation for it needs a placement-preview check (community offset +270). `D32` was not changed: it sits on the ring side, behind `Q5`, so a reversed supply never reaches it. Not covered: 12 V (or any wrong voltage) wired to a 5 V pin. Mainboard component cost per board is now $12.94 / $11.51 / $8.79 at 1 / 10 / 100 boards.
- **2026-10-17 `U2` is now a YLPTEC K7803-2000R3 (LCSC C18212378); `C12` and `C14` added.** ERC 0; DRC with schematic parity 0 violations, 0 unconnected, 0 parity issues; gerbers, BOM/CPL, ERC/DRC reports, schematic PDF and renders regenerated (kicad-cli 10.0.6); **not tested on hardware.** Why: the TSR 2-2433N was $9.27 with 4 in stock; the K7803 is about $2-3 with 607 in stock, same pin order (1 Vin, 2 GND, 3 Vout) and 2.54 mm pitch, so the pads and routing are unchanged. Its datasheet requires 22 uF at input and output: `C12` (+12V_MOTOR side, left of `U2`, 1 mm track to pin 1) and `C14` (right of `U2`, 0.5 mm track from pin 3), Samsung CL21A226MAYNNNE 25V X5R 0805 (C602037, 104,650 in stock), both on the top side so assembly stays single-sided; each ground end goes through a stub and a via to the B.Cu plane. `U2`'s body outline and courtyard were redrawn for the 11.5 x 9.0 mm body (the footprint name is still the TRACO one; the pads are identical). The module is 17.5 mm tall (was 10.2 mm) and has no input reverse protection (`D1` covers it). Narrowest +5V pour across the new copper stays at 11.6 mm. Mainboard component cost per board: $12.03 for one, $10.78 for ten, $8.23 for a hundred (was $17.90 / $16.87 / $14.62). The LCSC parameter table lists the K7803-2000R3 at 11.6 x 7.5 x 10.2 mm, which contradicts the maker's drawing; measure the first part. The `F1` gerber note above is resolved by this re-export.
- **2026-10-17 `R26` part changed to Yageo RC0603FR-0710KL (C98220).** C25804 (Uni-Royal, same 10 kΩ 1 % 0603) showed 0 in stock; C98220 has 3.4 million, $0.0036 at 100. Same footprint and value, so no board change. BOM CSVs and `build_jlc_assembly.py` updated. `U2` candidates (YLPTEC K7803-2000R3 C18212378 and others, not adopted yet) are in the README "Before ordering" list.
- **2026-10-17 mainboard BOM priced (tools and worksheet only).** New `kicad_mainboard/tools/bom_cost.py` prices `Mainboard_BOM_FULL.csv` from LCSC's product endpoint, picks the quantity break per part and checks stock; output in `jlc_export/mainboard_bom_cost.csv`. Component cost per board: $17.90 for one, $16.87 for ten, $14.62 for a hundred (2026-10-17 prices; excludes bare PCB, JLCPCB fees, fuses, harness and the plug-in modules). `U2` (TSR 2-2433N, $9.27 each) is 55% of it and had 4 in stock. `kickstarter/pricing_worksheet.xlsx` `MAINBOARD_BOM` now holds 14.62 (it was blank).
- **2026-10-17 harness build and test procedure (docs only).** New Hardware Overview section "Harness build and test (before first connection)": wire colour code and labels, an optional five-indicator jig (5 V good, 12 V good, 12 V on a 5 V pin, live spare pin, ground continuity), and a nine-step meter checklist from the bare harness through supply trim, fuse-by-fuse board power-up and the ring cable, plus WLED's maximum-current setting. Stopgap for the missing 5 V reverse-polarity and wrong-rail protection until a board revision.
- **2026-10-17 `F1` is now 2 A (was 2.5 A).** Littelfuse's ATO range (the datasheet kept locally) has no 2.5 A size (1, 2, 3, 4, 5, 7.5, 10, 15 A ...). Realistic worst-case draw on the 12 V side is about 0.9 A (motor ~0.45, `U2` ~0.3, fan 0.1), so 2 A holds with margin and keeps `D1` (1N5822, 3 A) inside its rating, where a 3 A fuse would not. Changed in the README, Hardware Overview, User Guide, BOM/cost CSVs, `build_jlc_assembly.py`, the schematic/PCB value text and the `F1` silkscreen text (PCB source only). **The gerbers in `kicad_mainboard/jlc_export/` still say "F1 2.5A ATO": re-export them before ordering** (kicad-cli was not available here). The Hardware Overview's 12 V supply sizing (about 2.5 A worst case) was left as a conservative figure.
** A+, A-, B+, B- silkscreen labels inside the `J3` housing outline, beside each pin (pin 1 = A+ at the bottom); the combined note was removed from the legend. DRC with parity 0/0/0; gerbers and top render regenerated; README silkscreen notes corrected (several reference designators are hidden on crowded parts).
- **2026-10-07 power-supply recommendation and wiring notes (docs only).** Hardware Overview section "Choosing a supply" and "Wiring the RD-85A to `J-PWR-IN`"; README suggests the Mean Well RD-85A (5 V / 12 V, 88 W), checked against its specification (no minimum load; 5 V 8 A rated, 10 A range). Prices unverified.
- **2026-10-11 LED ring board folded into the main docs (docs and reports only; board and firmware unchanged).** Hardware Overview section 6 now covers the ring as part of the project: size, LED pitch and chain, `J1`/`J5` cable table, power pours, full parts and LCSC table, assembly scope, component cost and ordering notes; sections 1, 9 and 10, the README ring bullet and ordering checklist, and the User Guide specifications updated to match. The ring's `erc_report.rpt` and `drc_report.rpt` were stale (the DRC still listed the `D89` jumper as unconnected, fixed on 2026-09-29) and were regenerated: ERC 0; DRC with schematic parity 0/0/0.
- **2026-10-09 OSC `/home`.** Recalls the stored home speed (`speedScale = homeScale`) and resets the emulated encoder position; unlike the button it never stops the platter when already at home. Compiles; not tested on hardware.
- **2026-10-09 main encoder button: home speed.** Short press (under 1 s) recalls the home speed (`speedScale`, signed, so it includes direction); pressing again while at home stops the platter, so the knob can still stop it (previously every press stopped it). Hold 1 s to under 5 s and release: the current speed becomes the new home (ignored while stopped). Hold 5 s: home returns to the factory value (1.0, full tempo speed forward), acted on as soon as 5 s is reached. Home is kept in flash (`homescale`), written while the motor is stopped like the other deferred settings, and reported as `homeScale` in `/status`. The button is now debounced for 30 ms on press and release. Compiles; not tested on hardware.
- **2026-10-07 board re-layout and re-route; pin map revised; `R24`, `R25`, `R28` removed.** ERC 0; DRC with schematic parity 0/0/0; firmware compiles; gerbers, BOM/CPL, schematic PDF and renders regenerated; **not tested on hardware, firmware not flashed.**
  - **Why:** the routing of the old placement was the main source of board complexity (about 8,100 on a score of copper length + 12 per via + segments; the new board is about 4,400, with fewer vias and a short, straight panel strip). The placement and pin map were optimised together by simulated annealing (straight-line length, crossings, barrier crossings) and routed with a purpose-built grid router (negotiated congestion, 0.1 mm grid, 0.2/0.2 mm rules); the scripts were scratch tools and are not in the repo.
  - **Layout:** the TMC2209 socket is rotated 90°, the panel headers are in the order MAIN, VAL, TAP, EXP3, EXP2, SAT, EXP1, HUE (left to right), and `D5`, `C8`, `C11`, `R27`, `J3` moved next to the motor pins.
  - **Pin map (`J1-H1`/`J1-H2`):** STEP 46, DIR 48, ENABLE 0, TMC UART TX/RX 47/15, WLED link TX/RX 18/2, I2C SDA/SCL 17/16, fan PWM 45, tap 21, main encoder CLK/DT/SW 38/39/40, hue CLK/DT 1/3, saturation CLK/DT 43/44, value CLK/DT 41/42. The main encoders stay on native GPIOs. Expander: hue SW `GPA0`, value SW `GPA5`, saturation SW `GPB5`, fan tach `GPA6`; expansion headers on the other channels (see README).
  - **Boot straps:** ENABLE on GPIO0 is held high by `R23` (also keeps the TMC2209 off through boot); STEP (46) and the fan gate (45, `R58` pull-down) are low at reset. `R24` (ENABLE pull-up), `R25` (DIR pull-down) and `R28` (GPIO43 series resistor) are no longer needed and were removed. GPIO21 (onboard WS2812) carries only the tap button; GPIO43/44 (boot-log UART) carry only encoder inputs.
  - **Firmware:** pin `#define`s updated; the expander is now read as one two-byte transaction (ports A and B, `mcp23017ReadPorts()`); comments and docs updated. Parts: 97 full-assembly parts / 392 joints (SMD 73 / 198), was 100 / 398.
  - **Antenna check:** copper under both antenna ends is still reported by `antenna_keepout_check.py`, but less than before (ESP32-S3-ETH end: no pads; DevKitC end: 10 tracks on F.Cu, no pads).
- **2026-10-06 wheel cartridge: bearing collars.** `enclosure/wheel_cartridge.scad` adds a 7.5 mm × 0.3 mm collar on the inside of each cheek around the axle, so the 625ZZ is held centred by its inner ring (0.1 mm axial play each side) instead of sliding 0.4 mm and rubbing its outer ring or tire on a cheek. New checks, rendered with OpenSCAD 2021.01: `check_bearing` (collars clear the outer ring and shields: empty) and `check_collar` (collars reach the inner-ring land: not empty); `check_wheel` and `check_mortise` still empty. `cartridge.stl` re-exported. `WHEEL_SUPPORT.md` now spells out the axle (press-fit 5 mm dowel pin through both cheeks) and the assembly order. Not printed or tested.
- **2026-10-06 motor rail 12 V; `U4`, `Q5`, `R19` removed; `D5` SMBJ15A.** ERC 0; DRC with schematic parity 0/0/0; gerbers, BOM/CPL, schematic PDF and renders regenerated; not tested on hardware.
  - **Why 12 V:** the 17HS08-1004S (3.7 Ω, 4.5 mH) at 0.8 A needs about 3 V for its resistance, and at this turntable's speeds (17 RPM default, even 100 RPM) inductance and back-EMF add under 2 V. 12 V gives the TMC2209 (29 V limit) about 17 V of margin instead of about 5 V, and lets `D5` clamp below every part on the rail.
  - **Circuit:** rails renamed `+24V_MOTOR*` to `+12V_MOTOR*` (schematic, board, the `Mainboard_power` rail symbol). `U4` and the optional `C12` pad are gone; `J-FAN` pin 2 is on `+12V_MOTOR`, joined on the board by a 0.5 mm F.Cu link from the old `U4` input to its old output track. `D5` is an SMBJ15A (GOODWORK C908796; alternate DIODES C135046): 16.7 V minimum breakdown, 24.4 V clamp, below the TMC2209 (29 V), `C8`/`C11` (35 V) and `U2` (36 V). `J-PWR-IN` pin 5 silkscreen reads *+12V*.
  - **5 V side:** `Q5` (reverse-polarity P-FET, about 1 W at full white) and `R19` removed at the user's request; `F3` kept. The fuse-output pour now overlaps the ring pour where `Q5` sat (at least 7.6 mm wide); the old `+5V_LED_RINGFUSE` net is part of `+5V_LED_RING`. Reverse polarity is now prevented only by the keyed connector.
  - **Parts:** 100 full-assembly parts / 398 joints (was 103 / 416). Docs (README, Hardware Overview, User Guide, power-flow diagram, jlc_export notes) updated for 12 V.
- **2026-10-06 stale schematic instance data fixed.** 54 symbols (`C6`, `C7`, `C10`, the connectors and others on the Power, Controller, Motor, LED and Panel IO sheets) carried a corrupted `(path "[(138, 64), ...]")` in their instances instead of the sheet path, left by an old generator. They now use `/<root>/<sheet uuid>`. That made KiCad see their no-connect pins, so the 55 no-connect pads on the board were given their `unconnected-(...)` nets to match: ERC 0; DRC with schematic parity 0/0/0. No copper, footprint or net connection changed; the gerbers differ only in net attributes. kicad-cli still prints "schematic has annotation errors" on export; the cause is not these paths.
- **2026-10-06 optional capacitor pads `C12` and `C13` (do not populate).** ERC 0; DRC with schematic parity 0/0/0; gerbers re-exported (same 12-file set); BOM/CPL unchanged because both are marked DNP. `C12` is a 0805 (10 µF/50 V X7R) across `U4`'s 24 V input, just below its pins at (9.0, 91.4); a 1210 does not fit there (courtyards of `U4` and `MH3`). `C13` is a 1210 (47 µF/6.3 V X5R) on the `+3V3` trunk at (60.5, 22.5), in the strip between the ESP32-S3-ETH and DevKitC sockets, with a ground via beside it. Not on the output of `U4` (no room without moving parts; the K7812 datasheet calls the output capacitor optional and the fan has its own). Schematic: new symbols on the LED and Power sheets with global-label / power-symbol connections. Purpose: an optional fix if the first-power-up scope check shows ripple on `+12V_FAN` input or `+3V3` (see the hardware review). Not tested on hardware.
- **2026-10-05 antenna keep-out check (tool only, board unchanged).** `kicad_mainboard/tools/antenna_keepout_check.py` lists the copper under the antenna end of each plug-in module (`--strict` exits 1 if any). Rectangles come from the Waveshare drawing (ceramic antenna / u.FL at the end opposite the RJ45) and an estimate for the DevKitC-1 (PCB antenna at the pin-1 end), and follow the socket positions. Current result: both ends have copper under them, as expected from the layout. ESP32-S3-ETH end: 22 + 11 track segments, 12 vias, `D5`/`R27`/`R28` pads, ground pour 68% on `B.Cu`. DevKitC end: 20 + 13 segments, 6 vias, the `U1` pins, ground pour 67%. Nothing is changed on the board; the check to make is Wi-Fi signal strength on the first assembled unit, and if it is poor, the ESP32-S3-ETH has a u.FL option (its antenna-select 0R resistor).
- **2026-10-05 `U4` BOM part switched to YLPTEC K7812-500R3 (C5369645).** $1.08 vs $8.05 and 1,598 vs 18 in stock, 36 V vs 32 V max input. Schematic, board and gerbers unchanged (same SIP-3 pinout, footprint still the TRACO outline); only the BOM rows and `tools/build_jlc_assembly.py` changed. Datasheet checked: same pinout and 11.50 x 7.55 x 10.20 mm body, so the footprint fits; the 22 uF/25 V input and output capacitors are optional (use 35 to 50 V parts if added). The `J1-H2` symbol/footprint value text no longer calls GPIO46 unused (it is the fan PWM pin).
- **2026-10-05 review follow-up (docs only).** Reworded the `D5` surge notes (README, Hardware Overview): the 42 V clamp is above the ratings of `U4` (32 V in), `C8`/`C11` (35 V) and `U2` (36 V), not only the TMC2209's. The tach note now covers a fan that pulls the tach to 12 V (about 1.8 mA through `R59`). `U4` fall-backs re-checked on LCSC: TRACO $8.05 with 18 in stock; YLPTEC K7812-500R3 $1.08 with 1,598; DEXU K7812-500R3 $0.95 with 1,016, both rated 15 to 36 V in. BOM part number unchanged; open question is whether the K7812 wants 22 uF in/out capacitors.
- **2026-10-05 fan header changed to a 12 V fan (cheaper fans), board still 100 x 100 mm.** ERC 0; DRC with schematic parity 0/0/0;
  the firmware compiles; not flashed or tested on hardware. A 5 V PWM 120 mm fan costs about 30 USD, a standard 12 V PWM fan 6 to 16 USD, so the
  header now supplies 12 V.
  - **Circuit:** `U4`, a TRACO TSR 0.5-24120 (24 V to 12 V, 0.5 A, short-circuit protected; same footprint family as `U2`), makes `+12V_FAN`
    from `+24V_MOTOR`. `F5` (the 5 V PTC) is gone: the converter limits its own output. `J-FAN` pin 2 is now +12 V; PWM, tach and the `GPIO46`
    inverter (`Q6`, `R57`, `R58`, `R59`) are unchanged. The fan runs whenever the 24 V supply is on, not only with the ring supply.
  - **Layout:** `U4` sits in the bottom-left corner under `J5` (courtyard x 3.2 to 16.3, y 81.2 to 90.3 mm). The header moved down 0.5 mm to clear it. The
    24 V feed is 53 mm of 0.4 mm copper on the bottom layer along x = 8.1 mm from `U2`'s input pin (one via near `U4`); it is kept off the front-layer
    +5V ring pour. It cuts a slit in the ground plane: ground return resistance went from 1.17 to 1.40 mΩ (18 to 21 mV at 15 A). The 5 V path is unchanged (1.43 mΩ).
  - **Safety:** a 5 V fan plugged into this header would be damaged, so the silkscreen reads *12V FAN* and the docs, web page and firmware comments say so.
  - **Parts:** `U4` C5993053 (TRACO, JLCPCB Extended, about 18 in stock; pin-compatible fall-backs YLPTEC K7812-500R3 C5369645, about 1.6k, and DEXU K7812-500R3 C909660,
    about 1.0k); `F5` and its part number are removed. Fan candidates: ARCTIC P12 PWM PST (12 V, 0.1 A, 200 to 1800 rpm) and Noctua NF-P12 redux-1700 PWM (12 V, 0.08 A).
    BOM: 103 full-assembly parts, 416 joints.
- **2026-10-05 fan defaults retuned, fan candidate checked.** Not yet flashed. On the bench a bare ESP32-S3-ETH idled at
  58.8 °C, so the shipped fan range moved from 50-70 °C to **60-78 °C** and the log warning from 75 to **80 °C** (the Arduino core
  configures the on-die sensor for its -10 to 80 °C range, so readings above that are less accurate). Saved settings still override
  the defaults. A 120 mm fan candidate (Coolerguys CG12025L05B2-4Y) is fixed-speed with no PWM or tach and does not fit the header; the
  Noctua NF-A12x25 5V PWM does (see README, Heat and ventilation).
- **2026-10-05 fan header added (100 x 100 mm board kept), with temperature control.** ERC 0 violations;
  DRC with schematic parity: 0 violations, 0 unconnected, 0 parity issues; the firmware compiles. Not
  tested on hardware, and the default temperatures are guesses until measured.
  - **Circuit** (`led.kicad_sch`): `J-FAN`, a keyed 4-pin 2.54 mm header (1 GND, 2 +5V, 3 tach not connected,
    4 PWM); `F5`, a 0.75 A PTC from `+5V_LED_RING`; `Q6`, a 2N7002 that pulls the fan's PWM input low; `R57`
    1k from `GPIO46` (`H2` pin 6, label `FAN_PWM` on the controller sheet) to the gate; `R58` 10k gate pull-down.
    Floating pin, reset, crash: FET off, fan at full speed. The strap pin sees the same low level as before.
  - **Layout:** all five parts in the bottom-left corner (header at y 93.5 mm, `F5` under `J5`'s +5V pad, `Q6`/`R57`/`R58`
    below the `J3` connector). `FAN_PWM` runs 80 mm from `H2` pin 6 on the bottom layer, 2 vias, kept out of the
    socket pad columns so the ground plane is not cut. Ground return resistance unchanged (1.17 mΩ).
  - **Firmware:** 25 kHz LEDC on `GPIO46`, inverted. Speed follows the smoothed chip temperature from a minimum
    (25%) at or below 50 °C to 100% at or above 70 °C, speeding up 10 points per 2 s and slowing 2 points per 2 s.
    New Fan box on the Motor tab and `POST /fan` (settings saved while the motor is stopped); `fanPct` in `/status`;
    the 5-minute log line includes the fan percentage.
  - **Fan tach and part numbers (same day).** The tach now reaches the expander: `J-FAN` pin 3 -> `R59` (4.7k, new, next to
    the fan parts) -> `U1` pin 3 (`GPB2`), 49 mm on the board with 8 vias, routed around the socket pad columns. DRC with
    schematic parity 0/0/0, ground return unchanged (1.17 mΩ). Firmware reads `GPB2` with the buttons (same I2C
    transaction), so it is a stall detector, not a tachometer: Status tab "tach turning / STOPPED / not detected",
    `fanTach` in `/status`, a log warning if a fan that was turning stops. Part numbers confirmed against LCSC/JLCPCB pages
    on 2026-10-05: `Q6` C8545, `F5` C976305 (BHFUSE BSMD1206-075-30V), `J-FAN` C164168 (Molex 22232041), `R58` C25744,
    `R59` C25900. BOM: 103 full-assembly parts, 415 joints.
- **2026-10-05 chip temperature monitor, heat estimates, fan-header proposal, wheel cartridge draft.**
  Firmware compiles; the temperature reading is not yet checked against a thermometer on hardware.
  - **Chip temperature.** The main ESP32-S3's on-die sensor is read every 2s from `loopMonitorTask()`
    (core 0), smoothed, and shown on the Status tab and in `/status` (`chipTempC`, `chipTempMaxC`).
    The log gets a line every 5 minutes and a warning above 75 °C. It tracks the module, not the
    ring board or the motor.
  - **WLED current limit** advice changed from about 8000mA to about 5500mA (README, hardware
    overview), pending a warm-room test of the enclosure.
  - **Heat estimates** for the 14 x 17 x 2 in plywood enclosure at 110 °F are in the README
    ([Heat and ventilation](README.md#heat-and-ventilation)): sealed it holds about 2 A of average ring
    current, vented about 3 A; high brightness for hours needs about 6 to 15 CFM.
  - **Fan header (proposed here; built in the entry above).** The board has no free GPIO except `GPIO46` (`H2` pin 6, a
    strapping pin; 33 to 37 are the module's PSRAM). Drive it through a 2N7002 N-FET (1k in series,
    10k gate pull-down) that pulls the fan's 4-wire PWM input low. With the pin floating or the firmware
    stopped the FET is off, the fan's own pull-up holds PWM high, and the fan runs at full speed, which is
    the safe failure. The strap sees the same low level it does today. 1x4 header: GND, +5V (from
    `+5V_LED_RING` through a 1A PTC), tach, PWM. Tach to the expander's spare input `GPA6` with a 10k pull-up
    to +3V3 for stall detection. Firmware: 25 kHz PWM, speed from the smoothed chip temperature between two
    settings, with a minimum speed and a full-speed kick at start. The only free F.Cu area big enough is
    the bottom-left corner (about x 9 to 16, y 81 to 95); `GPIO46` would route there on B.Cu, crossing
    the 3.3V trunk at x 28.8.
  - **Enclosure.** `enclosure/` holds a draft design for four platter support wheels (cartridge
    parametric OpenSCAD, rendered and interference-checked, not printed or tested).
- **2026-10-05 doubled the +3V3 trunk vias.** A copper-resistance check of the board showed the 3.3V
  trunk (about 1.25A peak) crossing layers through three single 0.3mm vias. Each now has a second
  one: (26.6, 64.3), (25.4, 62.1) with a short 0.5mm top-layer link to the trunk, and (29.7, 59.2)
  with a short 0.5mm bottom-layer link. Nothing else moved. DRC with schematic parity: 0
  violations, 0 unconnected, 0 parity issues. Gerbers, renders and the DRC report re-exported;
  BOM/CPL unchanged. Not yet tested on hardware.
- **2026-10-05 panel ESD protection; mainboard re-laid out from scratch; panel pin map changed.** Not
  yet tested on hardware. ERC 0 violations; DRC with schematic parity 0 violations, 0 unconnected,
  0 parity issues; the firmware compiles.
  - **Protection cells.** Each of the 21 panel signal pins (CLK/DT/SW of the encoders, `TAP_BTN`) now
    has a 1k 0402 series resistor and a PESD3V3S1UL TVS (SOD-882) to GND on the connector side
    (nets `NAME_J`). Each encoder header's `+3V3` pin is fed through a 100R 0402. `D32`, an SMBJ5.0A
    in SMB, sits across `+5V_LED_RING` and GND beside `J5`. 21 new TVS, 28 new resistors and one
    SMB diode in total.
  - **Room made by dropping `J-EXP` and `J-ENC-EXP4`.** The board stays 100 x 100mm, two layers.
    The panel headers moved to one row along the bottom edge (7.6mm pitch, order TAP, VAL, MAIN, SAT,
    HUE, EXP1, EXP2, EXP3), each with its cell beside it, and all 21 lines were routed again.
  - **Pins reassigned so the lines don't cross.** The firmware constants changed to match:
    main encoder A/B/button 1/2/43 (was 17/18/16), tap 21 (was 15), hue A/B 16/17 (was 1/2),
    saturation A/B 15/18 (was 21/43), value unchanged 47/38. The main encoder's button is now the
    one behind `R28` on GPIO43 (the boot-log pin). Expander buttons: hue `GPB3`, value `GPB0`,
    saturation `GPB1` (were `GPB0`/`GPB1`/`GPB3`). Existing panel wiring does not need to change
    on the module side, but the headers' signals land on different processor pins, so a board
    built from the old layout needs the old firmware. Every net on `J1-H1`/`J1-H2` was checked against the
    GPIO column of Waveshare's published schematic (H2 pin 5 = GPIO47, pin 14 = GPIO38, pin 4 = GPIO48,
    pins 9-12 = GPIO42/41/40/39, pin 7 = GPIO45; H1 pin 4 is `3V3_EN`, held low on purpose): all match the
    firmware constants.
  - **Silkscreen.** Each header pin has its letter again (V S D C G, `x` for the unconnected `J-ENC-EXP3` SW pin,
    G/T on `J-BTN-TAP`) to the right of the pin, with the one-line legend above the row. The first version of this
    layout dropped them; they were added back after review.
  - **Fabrication files regenerated:** gerbers, BOM/CPL (SMD 74 parts / 210 joints, full 97 / 400),
    schematic PDF, renders, ERC/DRC reports.
- **2026-10-04 faster loop-stall reset (3s), starting-colour picker, colour temperature modifier.**
  - **Loop monitor.** The IDF watchdog took about 12s to reset a hung `loop()`. `loop()` now bumps
    `loopBeat` each pass and `loopMonitorTask()` (core 0, 100ms poll) restarts the chip when it
    stops for 3s, after switching the driver off and stopping the step timer itself. A marker in
    RTC memory makes the next boot log `Reset reason: LOOP STALL`. The IDF watchdog stays as the
    backstop. Measured on the bench with the `WDT_TEST_HANG_MS` hook: about 3s instead of about 12s.
  - **Starting colour.** A colour picker on the settings page (`startcolor`, `#rrggbb`, persisted as
    `startrgb`, default red) sets what the light shows at switch-on. `setup()` converts it to
    hue/saturation/value (brightest channel = brightness). New `colour.h` holds the conversions
    (`hsvToRgb` moved out of the sketch unchanged, plus `rgbToHsv`, `parseHex`, `formatHex`), with
    `tests/test_colour.cpp` (53,000-colour round trip, hex rejects). Boot application checked on
    the bench with a temporary default (`#004080` gave `bri=128 rgb=0,128,255`); the web form
    itself is not yet tested on hardware.
  - **Colour temperature.** A number on the settings page (`ctemp`, 2000 to 10000 K, persisted,
    default 6500 = neutral). `colour::applyTemperature()` multiplies the RGB sent to WLED by
    per-channel gains from a black-body curve fit, normalised so the largest gain is 1 (it only
    ever dims a channel, so colours never go past full brightness). Warm leaves red alone and cuts
    blue; cool does the opposite. Applied after Save & Reboot; unit-tested, the settings form not
    yet tested on hardware.
  - **Test runner** now builds every `tests/test_*.cpp` (`run_tests.bat` / `run_tests.sh`).
  - **Colour temperature is live.** OSC `/colourtemp` (alias `/colortemp`; kelvin, float or int,
    clamped 2000 to 10000), a **Colour temperature** slider on the `/live` page (`/control?a=ctemp&v=`)
    and `colourTempK` in `/status`. The live value (`colourTempK`) is separate from the saved
    switch-on setting (`startColourTempK`), so changing it live never alters what the settings page
    saves; a restart returns to the saved value.
  - **Dark theme and tabbed settings page.** One shared stylesheet (`THEME_CSS`, dark only) now styles
    every page: config, live, log, update and the Saved / Invalid / Update-OK message pages
    (`PAGE_TOP`). The config page is split into tabs (Status, Network, Show, Colour, Motor,
    Security) inside the same single form, with a sticky **Save & Reboot** bar, the tab kept in the
    URL fragment, and an invalid field on a hidden tab switching to that tab. Without JavaScript all
    tabs show. The smooth-motion floor slider moved to the Motor tab (still live, still unnamed so
    the form never submits it). Compiles; not yet seen in a browser on the bench board (the pages
    are password-protected), so the layout is untested.
  - **Documents served by the unit.** `docs/build_docs.py` now also writes `src/docs_data.h`, the
    two HTML documents gzipped (about 128 KB of flash), and the unit serves them at
    `/docs/USER_GUIDE.html` and `/docs/HARDWARE_OVERVIEW.html` (`/docs` redirects), behind the usual
    password and linked from the config and live pages. The on-device copy is always dark, shows the
    links to repository files that are not on the unit (README, changelog, schematic PDF) as plain
    text, and has a Settings link in the top bar. `--check` (and CI) also fails if the header is
    out of date, by comparing the SHA-256 of each document recorded in it rather than the gzip
    bytes, which can differ between zlib versions.

- **2026-10-04 bench session: log lines no longer interleave, BPM freshness stamps moved inside the
  lock, watchdog proven.** First end-to-end run of the firmware on the bench board, using real OSC
  messages over Wi-Fi.
  - **What the bench confirmed (no change needed):** `/bpm` accepted and a tempo-controller message
    ignored while it was fresh; eight `/scaleEncoder` steps gave exactly +0.04 each; the start began
    at the 64 steps/s floor and settled at 441.7 against 441.8 from the formula; hue 0.5 gave
    `rgb=0,255,255` and half brightness gave `bri=128`; saturation 0 gave white; a reversal went
    441 to 139 to -441 without the driver stopping; `/scaleZero` ramped -441 to 0 in about 0.5s.
  - **Log interleaving (bug, found on the bench).** Messages from two tasks mixed mid-line
    (`BPM (OSC): BPM Source: OSC /bpm130.00`) every time an OSC message changed the tempo source.
    The old lock only kept each `write()` call whole, and a message is several calls. `LogTee` now
    keeps a partial line per task (12 slots) and writes each finished line in one piece. The prefix
    time is the moment the line completes. An audit found all 53 `Log.print(` calls end their line.
  - **Freshness stamps (ordering fix, not reproduced).** `onBpm()` and `onTempoController()` set
    `lastBpmReceived` / `lastTempoControllerReceived` and the "ever received" flags just after
    releasing `stateMux`, leaving a gap in which `loop()` could see the source with a stale stamp and
    demote it. They are now set inside the same critical section. I could not make the race happen in
    four trials; this is closing a theoretical window.
  - **Reset reason at boot.** Every boot logs why the chip last restarted (power-on, brown-out,
    crash, interrupt or task watchdog, firmware update, ...). After a watchdog reset the log is empty
    except for this line.
  - **Watchdog tested for real.** A temporary `WDT_TEST_HANG_MS` build made `loop()` spin at 20s of
    uptime. Result, twice: the ESP-IDF reported `loopTask (CPU 1)` not resetting the watchdog, the
    chip reset, and the next boot logged `Reset reason: TASK WATCHDOG`. **The reset came about 12s
    after the hang, not the configured 5s**, so the comments, log line and README now say about 12s.
    The hook stays in the source (compiled out unless `WDT_TEST_HANG_MS` is defined). Pings to the
    board stopped replying roughly 6s into the hang, before the reset; the cause was not investigated
    (the first test's single failed ping at +24s looked the same but one probe is not evidence).
  - **Seen, not fixed:** in two of four boot trials the board had not joined Wi-Fi within 14s, in line
    with the 5 to 25s joins seen on this network.
- **2026-10-04 review fixes: loop watchdog, shared-name race, hot-plug warning, CI.**
  - **Loop watchdog (safety).** The core's task watchdog watches only core 0's idle task and the
    step timer is an interrupt, so a hung `loop()` (core 1) would have left the platter turning at
    its last speed with the driver enabled and nothing resetting the chip. `setup()` now ends with
    `enableLoopWDT()`: a `loop()` that stops coming round resets the chip (about 12s, see the bench entry above), and `R24` holds
    the driver disabled until `setup()` has run again. The slowest legitimate pass is a networked-
    WLED HTTP push (about 0.6s). Boot log line: `Loop watchdog on: ...`. Not exercised on
    hardware (the bench board was off the bench for this change).
  - **Race on two shared names (bug).** The Wi-Fi network being tried (written by `wifiTask()`, read
    by the event handler) and the announced mDNS name (written by `mdnsTask()`, read by the web
    handlers) were Arduino `String`s. Reassigning a `String` frees its buffer, so a concurrent read
    could touch freed memory. Both are now fixed `char` buffers written and copied under a lock
    (`setSharedName()` / `sharedName()`).
  - **No-hot-plug warning.** The User Guide, Hardware Overview and README now say not to connect or
    disconnect the motor (`J3`), the ring cable (`J5`) or a plug-in module with power on: an
    unplugged motor makes a spike that can destroy the TMC2209. The quick start now says to check
    connections with the power off.
  - **README task list.** The dual-core section now has a table of every task (`loop()`, `oscTask()`,
    `wifiTask()`, `mdnsTask()`, the event task, the step ISR) and describes the watchdog.
  - **Repo hygiene.** `.gitattributes` (`*.sh` LF, `*.bat` CRLF; a `.bat` with Unix line endings can
    break `goto` labels). `python docs/build_docs.py --check` exits 1 if an HTML page is out of date
    (tested by tampering with a page). `.github/workflows/ci.yml` runs the unit tests, the docs check
    and a firmware compile on every push; **the workflow has not run yet**, so its first run on GitHub
    is its first test (its YAML parses, the pinned library versions exist in the index, the unit tests
    pass under gcc's parser, and a build without `wifi_defaults.h` was done locally: 1,164,367 bytes).
  - **Build-time note, revised.** Removing the top-level `wifi_defaults.h` made the next build a
    full library rebuild (13m40s); earlier full rebuilds also followed top-level files being added.
    Adding files under `docs/` did not. So the likely trigger is a change to the set of top-level
    files in the sketch folder, though it is still not proven.
- **2026-10-03 the SAT knob's push button toggles white** (was: saturation back to 100%).
  White is saturation 0%, which gives RGB 255,255,255 at the current brightness; the hue is left
  alone. Pressed with colour showing, it remembers the saturation and goes white; pressed again
  while white, it restores that saturation. If the light is already white because the knob was
  turned down to 0, a press goes to 100%. Turning the SAT knob forgets the remembered value, so a
  press can never jump to a stale setting. The saturation encoder's partial click is cleared on a
  press. The log says `White (saturation button; was 73%).` and `Saturation back to 73% (saturation
  button).` The HUE button is unchanged (hue back to red) and so is the VAL button (brightness back
  to 100%). Comments and the User Guide updated; the signal diagram's label "3 reset buttons" is now
  "3 push buttons". The buttons are on the panel header (expander `GPB0`/`GPB1`/`GPB3`), which the
  bench unit does not have, so the behaviour was not exercised on hardware.
- **2026-10-03 log page rebuilt for long logs.** The old page had its only control (a "follow"
  box) at the top of a page that scrolled away, redrew the whole log every 2s, and yanked the view
  to the bottom each time, so reading back was nearly impossible.
  - **Pinned top bar:** the title plus **Live: on/off**, **Clear**, **Save .txt**, **Config**, and
    a status (line count, last update, or "offline, retrying"). Live switches the refresh and the
    follow-the-end scrolling together; off freezes the page.
  - **Behaves with long logs:** follows the end only while you are already at the end, and does
    not touch the text when nothing has changed, so scrolling up and selecting text both work.
  - **Clear** is a new `POST /log/clear` (authorised, same-origin checked) that empties the unit's
    ring buffer; the first line afterwards says it was cleared. It clears the log on the unit, not
    just the view.
  - **"Raw" is now "Save .txt".** It was a link to `/log.txt`, the same log as plain text for
    copying or `curl`; it now downloads `turntable-log.txt`. `/log.txt` itself is unchanged.
  - The page is one fixed document in flash (like `/live`). Tested with a local stub of the log
    endpoints (layout in headless Edge) and 18 scripted checks of the page logic in Node (live
    toggle, clear, follow rules, no rewrite when unchanged, offline report).
- **2026-10-03 the two documents are now HTML by default.** `docs/USER_GUIDE.html` and
  `docs/HARDWARE_OVERVIEW.html` are self-contained pages (inline stylesheet with light and dark
  themes, a contents sidebar that folds into "On this page" on phones, print styles, the diagrams
  as inline SVG and the board picture embedded as a downscaled JPEG), so a single file can be
  e-mailed or copied. The README links to them first.
  - **Markdown stays the source.** `docs/build_docs.py` (about 430 lines, standard library only; it
    uses Pillow only to embed the photograph) rebuilds both pages from the `.md` files, so the two
    formats cannot drift apart. It handles the subset the documents use and was checked by
    comparing heading, table and list-item counts against the source, and by rendering the pages
    in headless Edge in dark and light themes and at phone width. GitHub shows `.html` as source,
    so the README also points to the Markdown for browsing there.
  - **Diagrams** are now `docs/diagrams/power-flow.svg` and `signal-flow.svg`, used by both
    formats, instead of Mermaid blocks that only rendered on GitHub. They follow the page theme.
- **2026-10-03 title header on every admin page.** "Uncle Magic's Liquid Turntable" now heads
  every HTML page the unit serves, as an `<h2>`: the settings page (it was an `<h1>`), `/live`,
  `/log`, `/update`, and the short result pages ("Saved. Rebooting", invalid input, "set a password
  first", "Update OK"). One `BRAND_H2` macro holds the markup for the small pages, which have no
  stylesheet. The update page's own "Firmware update" heading is now an `<h3>` so the levels stay
  in order. The plain-text and JSON replies (`/log.txt`, `/status`, `/control`) are not pages and
  are unchanged. Flashed to the bench unit; it boots and serves normally (the pages themselves sit
  behind the web password and were not viewed).
- **2026-10-03 hostname `liquidturntable`, editable, with automatic conflict numbering**
  (flashed to the bench unit and tested).
  - **Default name** is now `liquidturntable` (was `stepper-osc`): the DHCP and Wi-Fi hostname and
    the mDNS name (`liquidturntable.local`). Anything bookmarked or scripted against
    `stepper-osc.local` needs the new name.
  - **Editable** on the config page ("Name on the network"); validated as a DNS label (lower-case
    letters, digits, hyphens, 1 to 31 characters, no hyphen at either end), saved in flash, applied
    after the reboot. The config page, `/live` and `/status` (`"host"`) show the name in use.
  - **Conflict numbering.** `mdnsTask()` brings mDNS up under a temporary chip-unique name, asks the
    network whether the wanted name is taken, and if so tries `name-1` to `name-9`; the name is then
    announced with `mdns_hostname_set()`. A random 0 to 0.8s delay first keeps simultaneous
    power-ups from all finding it free. It runs as its own task so the network event task is not
    held up. Hyphens, not underscores: an underscore is not valid in a hostname and some systems
    refuse to resolve it.
  - **Tested** with a temporary build whose default name was this PC's own name: the log said
    `"tiny.local" is already used by 10.1.10.14`, the unit announced `tiny-1.local`, `tiny-1.local`
    reached the board, and `tiny.local` still reached the PC. The default was then restored and
    re-flashed.
  - **Wi-Fi fix found along the way:** an access point can take longer than the 15s attempt limit
    to hand out an address after a successful association (seen once at 20s). The attempt limit
    is now 15s to join plus up to 40s for the address, so a good connection isn't torn down while
    DHCP is still working.
  - Not changed: with several units, which one gets which number depends on power-up order; set
    distinct hostnames on the config pages for names that stay put.
- **2026-10-03 project renamed "Uncle Magic's Liquid Turntable"; user guide and hardware
  overview added.**
  - **Name** applied to the README title, the config, `/live` and `/log` web pages' titles, the
    first line of the boot log, the root schematic sheet's title (schematic PDF regenerated; ERC
    still 0) and the Kickstarter page. The GitHub repository name (`ESPMagicTurntable`) is
    deliberately unchanged, because renaming it breaks existing links. The network hostname
    changed from `stepper-osc` to `liquidturntable` (next entry).
  - **`docs/USER_GUIDE.md`**: for the person running it: quick start, the panel, speed and tempo,
    lighting, the phone and settings pages, OSC, network setup, power and safety, maintenance,
    troubleshooting and specifications.
  - **`docs/HARDWARE_OVERVIEW.md`**: for the person building or servicing it: system and
    power-flow diagrams, the mainboard and its modules, rails and protection, pin use, connector
    reference, the LED ring board, motor and mechanics, ordering notes, verification status.
  - Both documents are written from the design files and the firmware, and say plainly that no
    fully assembled unit has been tested yet.
- **2026-10-03 built-in Wi-Fi networks** (flashed to a bench ESP32-S3-ETH and confirmed
  working: the log shows `saimasnet` not found, `KW` connected, IP assigned, mDNS started, and the
  web page answered by IP and by `stepper-osc.local`).
  - **Candidate list.** Wi-Fi networks are tried in order: the one on the config page, then up
    to 4 built-in ones from `wifi_defaults.h`. That file is gitignored (the passwords stay out of
    the repo); `wifi_defaults.example.h` is the template. A new `wifiTask()` gives each network
    15s, moves on at once when one is rejected, and when none connects waits 15s and starts
    again from the first. Auto-reconnect is off, because the driver would retry an absent
    network forever and never reach the next one. Blank on the config page no longer means
    "Wi-Fi off" when built-in networks exist.
  - **Bug found on the bench and fixed:** the driver sends `STA_LEAVING` (reason 36) as an attempt
    starts and after every `NO_AP_FOUND`. It was counted as a failure, so a good attempt could be
    abandoned for the next network. Those events are now ignored.
  - **Build note (corrected twice).** Several builds took 10 to 16 minutes because arduino-cli
    rebuilt every library. See the 2026-10-04 entry for the current best explanation (a change to
    the top-level files in the sketch folder); adding files under `docs/` did not trigger one.
    Normal edits rebuild in 35 seconds to 3 minutes.
- **2026-10-03 first boot (DHCP + mDNS), gentle stop, `motion.h` with unit tests, README
  restructure** (firmware compiles; `tests/` passes; **not yet tested on hardware**).
  - **DHCP by default.** The firmware defaulted to a fixed address from the author's network
    (172.16.104.227, gateway .1) that the README never mentioned, so a new board on any other
    network was unreachable and invisible until someone read the USB log or did the tap-button
    reset. It now defaults to DHCP. A board that saved a static address keeps it; one that
    never saved any setting was on the old default and moves to DHCP.
  - **mDNS.** `http://stepper-osc.local/` answers on every link that has an address, and the
    config page shows the name. The README's "First boot" section lists the ways to find a board.
  - **Gentle stop.** Stop (encoder button, `/scaleZero`, `/live` Stop, the deadzone) used to cut
    the driver at once, even at full speed, while every start ramped. It now glides down to the
    smooth-motion floor (about 1.5s from full speed) and then disables the driver.
  - **`motion.h` and `tests/`.** The speed control law (BPM ramp, deadzone, floor remap, slew,
    reversal, stop) moved out of `loop()` into `motion::step()`, with no Arduino headers.
    `tests/test_motion.cpp` covers it, including a 400,000-step randomized run that checks the
    invariants: 1.17 million checks pass, and deliberately breaking the stop, the 50ms cap, the
    parked-motor handling or the hysteresis makes it fail. Run `tests/run_tests.bat` (Visual
    Studio Build Tools) or `tests/run_tests.sh` (g++/clang++).
  - **README rewritten as current-state documentation.** The dated history (nearly half of the
    file) moved here verbatim. "Known open items" became "Open items" (bench checks, checks
    before ordering, deliberate gaps) plus a new "Board notes". Stale statements fixed: the
    rollback section said Ethernet only; "1.25MB slots, image ~700KB" (now 3MB and ~1.1MB); a
    "known gap" about unlocked `speedScale` reads that was fixed long ago; the intro's mention of
    a diagram that no longer exists; the `H2` wired-pin count (12, not 13); the persisted-fields
    list now includes Wi-Fi and reverse motor direction.
  - **Hardware cross-check (no change needed).** Every ESP32-S3-ETH header pin in the netlist,
    and every GPIO number in the firmware, was checked against Waveshare's pinout diagram, and
    the TMC2209 headers against BigTreeTech's pinout table; both agree. The board also meets
    JLCPCB's two-layer limits (smallest drill 0.3mm, smallest annular ring 0.15mm, tracks 0.2mm
    and up).
- **2026-10-02 small fixes.** The config page's Wi-Fi password field now takes 64 characters, so
  a 64-hex-digit WPA key (which the firmware already accepted) can be entered. README and
  firmware notes: press the ESP32 module's BOOT button only with the motor stopped (it shorts
  GPIO0, the STEP output, to GND).
- **2026-10-02 ramp and step-timer hardening; motor direction setting** (firmware compiles;
  **not yet tested on hardware**).
  - **Ramp gaps closed.** After a failed or stalled firmware upload, the motor resumed at its
    old speed with no ramp, because the upload parks it without going through the ramp code.
    It now ramps up from rest. One slow `loop()` pass could also become a single big speed jump
    (up to ~7 RPM on the WLED HTTP path); the ramp's time step is now capped at 50ms.
  - **Step timer.** When the step period shrinks (ramping up, or restarting), the counter can
    already be past the new alarm value. The IDF doesn't say whether the ESP32-S3 then fires at
    once or only after the 54-bit counter wraps (effectively never: the motor would silently
    stop stepping). The firmware now resets the count in that case and before every start.
    Also, a restart within 10ms of the last period change used to run briefly at the old rate;
    a (re)start now always applies the new period.
  - **Reverse motor direction** (config page, saved): flips DIR for every move, for a motor
    whose coil is wired the other way round. `/directionChange` still toggles at run time but
    isn't saved.
  - README: footprint count in the PCB notes corrected (53, not 37).
- **2026-10-01 `J3` re-pinned to the standard stepper order.** `J3` used to pair pins 2+3 and 1+4
  (B1, A2, A1, B2), so a standard stepper cable made the motor buzz instead of turning. It is
  now 1 A+ (`MOTOR_COIL_A1`), 2 A− (`A2`), 3 B+ (`B1`), 4 B− (`B2`), matching StepperOnline's
  17HS08-1004S (black, green, red, blue) and ready-made printer cables. Only pins 1 and 3
  changed net.
  - **PCB:** A1 now drops to a via (0.8/0.4mm) above pin 3 and reaches pin 1 on B.Cu, passing
    above pin 2. B1 runs straight from its diagonal onto pin 3 at 0.6mm. That removes B1's old
    0.25mm dog-leg under pins 2–3. The pin labels on the silkscreen now read A+ A− B+ B−.
  - **Checks:** ERC **0**. DRC with schematic parity: **0 violations, 0 unconnected, 0 parity
    issues**. Gerber diff: only F.Cu, B.Cu, F.Silkscreen and one via in the drill file changed.
    BOM/CPL unchanged. Schematic PDF, ERC/DRC reports and renders re-exported.
  - The schematic generator now numbers UUIDs per sheet, so a change on one sheet no longer
    renumbers the others. This commit renumbers every sheet once.
- **2026-10-01 review fixes: speed slewing, `R28` on GPIO43, schematic tidy-up**
  (firmware compiles; **not yet tested on hardware**).
  - **Every speed change now ramps.** `/live` Start used to jump the platter from rest to full
    tempo speed in one step, and Faster/Slower and big `/scaleEncoder` moves stepped instantly
    too; only BPM changes and reversals ramped. A direct-coupled platter can stall or skip on
    that. The final step rate is now slewed at the BPM ramp's rate for every change (see
    [Speed ramping](README.md#speed-ramping-and-the-smooth-motion-floor)). A reversal no longer switches
    the driver off for a moment at the zero crossing. Stopping still cuts the driver at once.
  - **`R28`, 1k 0603 in series with `SAT_DT`** (GPIO43, the saturation encoder's DT). GPIO43 is
    UART0 TX, and the ESP32-S3 ROM drives it with its boot log at every reset, so an encoder
    resting with its DT contact closed would short that output to GND. The encoder side is the
    new net `SAT_DT_ENC`. On the PCB, `R28` sits just left of `J1-H1` pin 20 (under the module),
    feeding the existing B.Cu trace through one via. YAGEO RC0603FR-071KL, LCSC C22548, the same
    part as `R9`/`R23`/`R25`.
  - **Schematic:** the "Panel I/O" sheet is now "Panel IO" (KiCad uses `/` as the
    sheet-path separator). `U1` now uses KiCad 10's `MCP23017x-x-SP` library symbol, so the
    old "symbol not found in library" ERC warning is gone. The strap-resistor note moved from
    the LED sheet to the Controller sheet. A stale note saying F2/D2 were "not assembled" (they
    were removed) is corrected. Rotated parts' reference/value text now reads horizontally.
  - **Checks:** ERC **0**. DRC with schematic parity: **0 violations, 0 unconnected, 0 parity
    issues**. Gerber diff against the previous export: only F.Cu/B.Cu, the F.Mask/F.Paste
    openings for `R28`, the F.Silkscreen label and the via in the drill file changed. BOM/CPL:
    **24 SMD parts / 68 joints, 49 full-assembly / 267 joints.** Schematic PDF, ERC/DRC reports
    and renders re-exported.
  - **`J3` coil pairing** is unchanged (see the bench checks below): re-pin a standard printer
    stepper cable rather than re-routing the board.
- **2026-10-01 schematic split into hierarchical sheets.** The flat schematic was one oversized
  page. Every connection was a global label at a pin, with no wires, so you couldn't follow a
  signal by eye. It's now a root index plus five A3 function sheets (see
  [Hardware](README.md#hardware)), with wired connections, power symbols for rails, and global labels only
  between sheets. The netlist is unchanged: every pin connects to the same pins under the same
  net name as before. Symbols keep their UUIDs. All 48 PCB footprints now carry their
  sheet paths (only 11 did before), so cross-probing works. The copper is untouched, and all 21
  Gerber layers are identical to the previous export. ERC: only the existing
  `MCP23017_SP not found in Interface_Expansion` library note (the original's 170 off-grid
  warnings are gone). DRC with schematic parity: **0 violations, 0 unconnected, 0 parity issues** (the previously
  committed report still listed 81 parity notes).
  Schematic PDF and ERC/DRC reports re-exported.
- **2026-09-30 tempo sources can no longer start the platter; `/live` control page** (compiles;
  **not yet tested on hardware**).
  - **Bug fixed:** a tempo "takeover" used to reset `speedScale` to 1.0. Senders that only
    transmit on change go stale between changes, so the next change restarted a platter someone
    had stopped (or undid a slowed trim). The tap button did the same on every tap. Tempo
    sources now set BPM only; see [BPM source failover](README.md#bpm-source-failover).
  - **New `/live`, `/status`, `/control`** (see [Web config API](README.md#web-config-api)): live BPM,
    source, RPM, a speed-cap warning, start/stop/trim/reverse/tap/set-BPM and HSV sliders from a
    phone.
  - **`/tapbpm`:** each tapped tempo is now broadcast over OSC alongside `/tap`.
  - Ableton Link was evaluated (it compiles against this core: C++ exceptions on, RTTI off via
    `ASIO_NO_TYPEID`) but is on hold for a licensing decision: Link is GPLv2+ unless Ableton
    grants a proprietary license, which would make the whole firmware GPL.
- **2026-09-30 fuse ratings silkscreened** between the clip pairs ("F1 2.5A ATO",
  "F3 15A ATO"). Both take the same ATO size, so the board now says which goes where.
- **2026-09-30 blade-fuse clips re-spaced (`F1`, `F3`).** Checked against XFCN's T508VS
  drawing and Littelfuse's ATO datasheet: the drawing's L = 19.04mm is between the *outer
  edges* of the clips' single pads, but the board used it between the clip origins. That put
  the grips ~13.9mm apart for a fuse whose blades are 9.3mm apart, so no ATO fuse would have
  fit. All four clips moved 2.28mm inward. Use standard or low-profile ATO fuses (a mini fuse
  needs a different layout). DRC clean; CPL and gerbers regenerated.
- **2026-09-30 firmware fixes from a fresh review** (compiles; not yet tested on hardware):
  - **Wi-Fi hostname** is now set before `WiFi.mode()`. In core 3.x `mode()` creates the
    station interface and takes its name then, so the old order could leave the board on the
    network under a default ESP32 name instead of `stepper-osc`.
  - **Wi-Fi connection failures are logged** once, with the reason (for example `AUTH_FAIL` or
    `4WAY_HANDSHAKE_TIMEOUT` for a wrong password, `NO_AP_FOUND` for a network name not seen).
    Before, a bad password showed only "connecting to …".
  - **`/setmin` saves are deferred until the motor is stopped.** The step-timer interrupt isn't
    IRAM-safe, so the core holds it off during a flash write; saving mid-motion caused a
    few-ms stall. `loop()` now writes the value once the step timer is off.
- **2026-09-29 `U2` → TSR 2-2433N (2A), 3MB OTA partitions.**
  - **`U2`: TRACO TSR 2-2433N** (2A, SIP-3 14×7.6mm, Vin/GND/Vout, [JLC C45406999](https://jlcpcb.com/partdetail/TRACOPOWER-TSR_22433N/C45406999)),
    replacing the 1A TSR 1-2433, for headroom now that both ESP32-S3s can run Wi-Fi. Same pads
    and traces, on KiCad's own `Converter_DCDC_TRACO_TSR2-24xxN_TSR2-24xxxN_THT` footprint
    (the TO-220 stand-in footprint is gone). The body extends south of the pin row. The plain
    TSR 2-2433 (LCSC C5357116) was rejected: its centred body collides with `D1`. DRC with parity:
    0 violations, 0 unconnected, 0 footprint errors.
  - **Capacitors (checked 2026-09-30 against the TSR 2N datasheet):** no external input or
    output capacitor is required; see [Power rails](README.md#power-rails). Check `U2`'s orientation in
    JLC's THT placement preview (rotation changed to 90°).
  - **Partition scheme** switched to `app3M_fat9M_16MB` (two 3MB OTA slots) in every build
    command, plus `sketch.yaml`. NVS stays at the same offset, so saved settings survive.
- **2026-09-29 OSC over Wi-Fi as well as Ethernet** (compiles; **not yet tested on
  hardware**).
  - **Config page:** new Wi-Fi network name and password fields (blank name = off; blank
    password keeps the saved one; an "open network" box clears it). The status line shows both
    links' addresses and the Wi-Fi signal.
  - **Both links:** OSC, `/log`, `/update` and the config page answer on both. OSC is pumped
    while either link is up; the subscriptions are made once, by whichever link comes up first.
    OTA rollback confirmation accepts either link.
  - **`/tap`** goes out as a subnet broadcast on each connected link (a bare
    255.255.255.255 leaves through only one interface).
  - **Radio settings:** modem sleep off (OSC latency), transmit power capped at 15dBm (+3V3
    budget, see Power rails).
  - Firmware grows to ~1.11MB, which is why the project moved to 3MB OTA slots (next entry up).
- **2026-09-29 review follow-ups: WLED resend, DevKitC pins, ring jumper, parity.**
  - **WLED state resend (firmware):** the colour/brightness used to go to WLED once at boot and
    then only on changes, so if WLED (slower to boot) missed the first push, the ring kept its
    WLED boot look until a knob moved. The serial path now re-sends the current state every 2s
    for the first 30s, then every 10s (not logged). The main board owns colour and brightness:
    a change made in WLED's own UI is overwritten at the next resend.
  - **DevKitC-1 power/ground pins:** only one GND (`J3` pin 1) and one 3V3 (`J1` pin 1)
    were wired. Per Espressif's pin table, `J1` pin 22 and `J3` pins 21–22 are also GND and
    `J1` pin 2 is also 3V3; those socket pins were on the board but unconnected. Now tied
    (`J6-J1B.2`, `J6-J3B.1/.2` to the B.Cu GND plane; `J6-J1.2` to `J6-J1.1` with a short
    trace). `J1` pin 21, the DevKitC's 5V, stays unconnected.
  - **LED ring:** the D89→D90 data link is now a routed trace around the centre hole, replacing
    the hand-soldered jumper wire (see `kicad_led_ring_offset/jlc_export/README.md`).
  - **Schematic parity cleaned up on both boards:** the PCB footprints carried the footprint name
    as their Value and lacked the library prefix on their IDs, and the ring's net names lacked
    the schematic's `/` prefix. That produced 81 (mainboard) and 404 (ring) parity notes, burying
    anything real. All are fixed, and mounting holes are marked board-only. With real library IDs,
    KiCad then flagged footprints that differ from the KiCad 10 library copy. A pad-by-pad check
    showed every footprint's copper is identical (silkscreen/text only), so both projects now
    ignore `lib_footprint_mismatch`. DRC with parity on both boards: **0 violations, 0
    unconnected, 0 footprint errors**. ERC: 0 errors.
  - Comments: the WLED-link comments now describe the real DevKitC-1 setup, and the encoder
    ISR/OSC race comment is corrected (a count can be lost, nothing tears).
  - BOM/CPL unchanged on both boards; gerbers re-exported for both.
- **2026-09-29 firmware review fixes** (compiles; **not yet tested on hardware**):
  - **Encoder step size:** the ISRs count 4 edges per KY-040 click, so every
    knob was 4× too coarse. `takeDetents()` now converts counts to whole
    detents (`encoderCountsPerDetent = 4`) and carries partial clicks over.
    OSC `/scaleEncoder` adds 4 counts per emulated detent, so knob and OSC
    agree. The hue/sat/val loops also read the ISR position once instead of
    twice (a count arriving in between was lost).
  - **Reversal ramp:** `/directionChange` at speed decelerated to a stop and
    then jumped straight to full speed the other way. It now climbs back at
    the same slew rate.
  - **MCP23017 port A pull-ups** (`GPPUA = 0x7F`) are now enabled.
  - **`/setmin` is POST with a same-origin check** (was an unprotected GET that
    could also write flash).
  - README: `D4` is described as ESD-only.
- **2026-09-29 `R27`, 4.7kΩ bleeder across `C8`/`C11`.** UNI-ROYAL 1206W4F4701T5E
  (LCSC C17936, 250mW), between `+24V_MOTOR` and GND, on F.Cu at (46.0, 57.4) just
  southwest of `D5`, clear of the B.Cu signal tracks at x=43.6 and x=47.19. Pad 1
  spurs onto the 1mm +24V trace at y=55.5. Pad 2 goes 0.5mm east to a 0.8/0.4mm GND via
  at (49.2, 57.4). Its silkscreen reference is hidden (no room, same as `C8`/`C11`).
  Value: 4.7kΩ, not the 1–2kΩ first floated. It draws 5mA / 0.12W (24V) or 0.14W (26V)
  in a 250mW package, and the discharge time constant with 200µF is ~0.9s.
  See [Mains input and power-on](README.md#mains-input-and-power-on).
  - DRC: 0 violations, 0 unconnected, parity issues unchanged (81). ERC: 0 errors,
    three more cosmetic off-grid warnings. BOM/CPL: **23 SMD parts / 66 joints, 48
    full-assembly / 265 joints**. Gerbers, schematic PDF, reports, renders re-exported.
- **2026-09-29 `R26`, 10k pull-down on `LED_DATA` at `U3`'s input.** Nothing held
  `LED_DATA` low before the WLED module booted: `R18` is only a series resistor and
  the DevKitC's GPIO4 floats for the first second or two, so `U3` could pass noise to
  the 178 SK6812s and flash the ring at power-on. `R26` (0603, LCSC C25804, same part as
  `R24`) sits just east of `U3` on F.Cu, its pad 1 tapped onto the `LED_DATA` trace
  and its pad 2 through a 0.6/0.3mm via to the B.Cu GND plane. It adds ~0.33mA when the
  data line is high, which the GPIO drives easily through `R18`. See
  [Mains input and power-on](README.md#mains-input-and-power-on).
  - DRC: 0 violations, 0 unconnected, parity issues unchanged (81, cosmetic). ERC:
    0 errors; three new cosmetic off-grid warnings on `R26`'s pin and wire ends, like
    the other resistors. BOM/CPL: **22 SMD parts / 64 joints, 47 full-assembly / 263
    joints**. Gerbers, schematic PDF, ERC/DRC reports and renders re-exported.
  - **Still to do:** in WLED, disable "Turn LEDs on after power up".
- **2026-09-28 `D5` TVS and second bulk cap `C11` on `+24V_MOTOR`; `F4` removed outright.**
  - **`D5`, SMBJ26A (LCSC C123820, SMB),** cathode on `+24V_MOTOR`, anode on GND, on
    F.Cu at (45.6, 53.5) under the ETH module, tapped straight onto the 1mm +24V trace
    (via two 1.0/0.5mm GND vias to the B.Cu plane). It starts conducting near 29V and clamps
    around 42V. Reason: the TMC2209's VS absolute maximum is 29V and hot-plugging the supply
    into an undamped LC (cable + `C8`) can ring well above 24V. Note this limits, not
    eliminates, a spike: the clamp level is still above the driver's 29V. **Also:** with the
    planned single IEC inlet the supply cable is permanent and the supplies ramp up on their
    own, so just never plug or unplug `J-PWR-IN` with mains on; and a 19–20V supply gives far
    more margin than 24V.
  - **`C11`, a second 100µF/35V (same part as `C8`, LCSC C176683)**, in the strip between
    the two modules directly north of `C8`, rotated 90° (pads vertical). `C8` already uses
    the largest 35V part that fits the 6.3×7.7mm space (a 220µF/35V is 8–10mm), so two in
    parallel give 200µF and half the ESR. `C11`'s + pad runs 1mm along the y=54.3 corridor to
    the 24V trace end; its − pad goes to two GND vias. Making room:
    - `I2C_SDA`, `WLED_LINK_RX` and `WLED_LINK_TX` were re-routed to pass through the gap
      between `C11`'s pads / around its east side.
    - `R23` (GPIO0/STEP pull-up) moved north beside `R24`, next to the +3V3 column. Its STEP
      link now hops to B.Cu through a via and reaches `J1-H1` pin 17 on B.Cu.
    - `C8`'s and `C11`'s silkscreen reference designators are hidden (no room). They are in the
      BOM/CPL.
  - **`F4` (2A PTC) removed, `+3V3_PSU` merged into `+3V3`:** `U2`'s output now lands directly
    on the +3V3 copper (U2 → +3V3 trace, no part between). `U2` is a 1A short-circuit-protected
    converter, so a 2A PTC could never trip on overload; it only added series resistance, and
    its hold current drops with temperature. In the schematic the redundant `PWR_FLAG` on
    `+3V3` was deleted, since `U2`'s power-output pin now drives that net.
  - **Check at order time:** `C11` is placed at 90° — check its polarity/rotation in JLCPCB's
    placement preview (+ pad is the lower one, next to `C8`).
  - DRC: 0 violations, 0 unconnected. ERC: 0 errors. BOM/CPL: **21 SMD parts / 62 joints, 46
    full-assembly / 261 joints**. Gerbers, schematic PDF, ERC/DRC reports and renders
    re-exported.
- **2026-09-28 LED data buffer and boot-strap resistors.**
  - **`U3`, a 74AHCT1G125 (Nexperia 74AHCT1G125GW, LCSC C12495, SOT-353),
    plus `C10` 0.1µF.** Powered from `+5V_LED_RING`, it sits beside `J5`:
    `LED_DATA` (3.3V, through `R18`) goes in, and `LED_DATA_5V` goes out to
    `J5` pin 1. It has to be the **AHCT** part, whose TTL-level inputs read
    "high" from 2.0V. The AHC version (C7468) needs 3.85V at a 5V supply and
    would not fix anything.
  - **ENABLE ↔ DIR swapped** between GPIO45 and GPIO3, with `R23` (1k STEP
    pull-up), `R24` (10k ENABLE pull-up) and `R25` (1k DIR pull-down). See the
    strapping note in [Firmware pins](README.md#firmware-pinout). ENABLE and DIR were
    re-routed with `tools/microroute.py`.
  - **Firmware:** `dirPin` 3 → 45 and `enablePin` 45 → 3. `setup()` preloads
    ENN's output latch before `pinMode()`, so the driver can't glitch on.
  - **Also deleted:** a redundant 0.2mm F.Cu GND trace from `J5` pin 3 to
    `J2-MODE` pin 3 (both are through-hole on the B.Cu GND plane). It crossed
    the spot `U3` now uses.
  - **Motor coil traces widened** (they were 0.2mm for up to 1.1A peak):
    now 0.6mm on most of the run. DRC stepped the necks between header pins
    down to 0.2–0.4mm, at most ~5mm per net.
  - BOM/CPL: **20 SMD parts / 60 joints, 45 full-assembly / 259 joints**.
    ERC: 0 errors. DRC: 0 violations, 0 unconnected, parity notes unchanged
    (cosmetic).
- **2026-09-28 bulk caps, 24V vias, +3V3 widths.**
  - **C6:** the 220µF bulk cap on `+5V_LED_PSU` was fed by a 0.2mm B.Cu trace
    through a via. The pour now covers it directly, and the trace is gone.
  - **C8:** the TMC2209 VM bulk cap had a 10mm, 0.4mm GND path to its only
    via. It now has a GND via at the pad.
  - **24V hop vias:** 0.8/0.4 → 1.0/0.5mm.
  - **+3V3 traces:** widened as far as DRC allows (mostly 0.5mm), cutting the
    regulator-to-ETH-module path from ~145 to ~80mΩ before F4's PTC.
  - **Q5 heat checked:** SQJ407EP R<sub>DS(on)</sub> is ≤7.1mΩ at
    V<sub>GS</sub> = −4.5V. At a full-white 10.7A that's ~0.8–1.1W, a
    ~55–80°C rise at R<sub>thJA</sub> 68°C/W. Fine against its 175°C limit.
- **2026-09-28 +5V pour widened at `J-PWR-IN` (was a ~1.3mm neck).** A
  narrowest-copper check found the pour still pinched right below the power
  connector's pad row, where the ring's 10.7A enters. The +24V trace from
  pin 5 to `F1A` crossed under the row on F.Cu, and an F.Cu GND trace looped
  pin 6 → pin 8 around NC pin 7. The only way through was a ~1.3mm gap west
  of pin 5.
  - **`F1A`/`F1B` swapped places.** They're identical clips, and the schematic
    is unchanged. The 24V-in clip `F1A` now sits west of the pour, and pin 5
    reaches it on B.Cu through a 1.0/0.5mm via. The fused clip `F1B` (east)
    feeds `D1` through a short spur inside the pour.
  - The redundant F.Cu GND loop is deleted. Pins 6 and 8 are through-hole on
    the B.Cu GND plane. The +5V zone is widened west to x=11 and east to
    x=34.5.
  - **Result:** the narrowest single path is 3.3mm, and total copper across
    the pad row is 10.8mm. Elsewhere on the ring path, the worst
    cross-sections are ≥5.5mm at `Q5`'s source leads and ≥16mm of B.Cu GND
    on the return. DRC: 0 violations, 0 unconnected.
- **2026-09-28 F2/D2 removed; the LED ring's +5V feed was a 0.2mm trace.**
  - **F2/D2 removed** from the schematic and PCB. They were confirmed dead
    first: `+5V_LED` reached only D2's cathode, `+5V_LED_F2` only linked F2
    to D2, and neither the firmware nor any other rail used them. The netlist
    after removal is exactly the old one minus those two parts and nets.
  - **Found while checking that area:** the +24V leg from U2 to the TMC2209
    ran across the `+5V_LED_PSU` pour on F.Cu. The original 0.2mm trace and
    the 1.0mm reroute took the same path. That cut the pour into two halves,
    `J-PWR-IN` side and F3 side, and the only link between them was a 0.2mm
    B.Cu trace, ~43mm long. So the ring's whole 10.7A feed would have run
    through a trace that fuses at ~4A. **Fixed:** the +24V now hops under the
    pour on B.Cu (1.0mm, between two 0.8/0.4mm vias at x=15.2 and x=27.9),
    which opens an ~11mm F.Cu path for the +5V. The thin B.Cu trace is
    deleted. Verified by deleting all B.Cu +5V copper on a scratch copy: the
    two halves stay connected on F.Cu alone. The GND return detours around
    the hop's west end through ~12mm of plane. DRC: 0 violations,
    0 unconnected.
- **2026-09-28 module sockets un-mirrored, 24V widened, TMC GND, MCP23017
  erratum** (scripted, in order: `widen_power.py`, `reroute_24v.py`,
  `swap_module_sockets.py`, `route_sockets.py prep` + Freerouting 2.4.1 +
  `route_sockets.py import`, `microroute.py` for the last few connections,
  `cleanup_after_route.py`; DRC afterward: 0 violations, 0 unconnected,
  schematic parity unchanged):
  - **Both plug-in modules were mirrored.** Checked against Waveshare's pinout
    and board photos and Espressif's DevKitC-1 pin layout:
    - With the ETH module RJ45-side-up, its VBUS column lands on the right,
      but the board had `J1-H1` (VBUS…) on the left.
    - The DevKitC-1 (component side up, USB at our edge) needs `J1` on the
      right, but the board had it on the left. An earlier fix rotated each
      socket 180° in place instead of rotating the pair.

    Plugged in, the DevKitC would have shorted +3V3 to GND, and every ETH pin
    was swapped. Fixed by swapping the x positions within each socket pair, so
    the modules sit exactly where they did. Every trace to those ~35 pins was
    re-routed, with temporary keep-outs so the router couldn't touch the
    high-current pours or the B.Cu GND return under them. A pin-by-pin readback
    of the finished board matches both vendors' drawings. See "Which way up
    the modules go" above.
  - **+24V widened.** The whole chain was 0.2mm, which would fuse (~4A/1s on
    2oz) before F1's 2.5A blade fuse opens. Now 1.0mm, with a 0.5mm floor in
    a few tight spots (~10A/1s). The three spots that hugged a pad (J-PWR-IN
    pin 6, U2 pin 2, F2) were re-routed, and the leg to the TMC2209's VM pin
    now runs 1.0mm across the top of J2.
  - **TMC2209 GND tie widened** (J2-MOTOR pin 2 → J2-MODE pin 2 → C6, 0.8mm
    where it fits). A 20mm F.Cu GND trace from U2 across the +5V input pour
    was deleted; U2's GND pin is on the solid B.Cu plane.
  - **Firmware: MCP23017 GPA7/GPB7 are outputs now** (Microchip's current
    datasheet: as inputs they can corrupt SDA during reads). `IODIRA/IODIRB =
    0x7F`.
  - **Known trade-off:** the new routing leaves little B.Cu ground under the
    ETH module and J2. Only slow signals run there, and the high-current
    return corridor down the left side is untouched.
  - The old `build_pcb_100.py` / `route_pcb_100.py` / `place_check.py` flow
    was removed: re-running it would re-create the mirrored placement and the
    old pour shapes. Use `tools/route_sockets.py` + `tools/microroute.py` for
    any future re-routing.

- **2026-09-28 Q5/Q6 pin mapping + LED-ring current path** (scripted:
  `kicad_mainboard/tools/fix_q5_q6_sch.py` then `fix_q5_q6_pcb.py`; DRC
  afterward: 0 violations, 0 unconnected, schematic parity only lost the
  removed parts' entries):
  - **Q5/Q6 were wired to the wrong pins.** The generic `Q_PMOS_GSD` symbol
    (1=G, 2=S, 3=D) sat on KiCad's `PowerPAK_SO-8L_Single` footprint, whose
    pads follow the physical pinout (1–3 source, 4 gate, tab drain). So the
    gate and drain nets landed on source pins, and the real gate and drain
    were unconnected. The FETs did nothing, and the ring's 10.7A would have
    run through two 0.7×0.4mm source pins. DRC can't catch this
    (unconnected pads are legal). Fixed with a project-local footprint,
    `lib.pretty/PowerPAK_SO-8L_Single_GSD`, whose pads are renumbered to
    match the symbol. Q5 keeps its position; its drain tab now sits in the
    `+5V_LED_RINGFUSE` pour and its source pins in the `+5V_LED_RING` pour,
    with the gate going straight to `R19`.
  - **Q6 and R20 removed.** Q6 only followed `U2`, whose output can't
    reverse (`D1` already blocks a reversed 24V input), and it sat in the
    middle of the ring's current corridor. `F4` now feeds `+3V3` directly.
  - **Ring pour un-cut.** `LED_DATA`, a redundant F.Cu GND trace between
    `J5`'s GND pins, and Q6's traces all crossed the `+5V_LED_RING` pour,
    leaving ~1–1.5mm necks (plus a 0.2mm B.Cu trace that DRC counted as the
    connection). `LED_DATA` now runs on B.Cu inside the pour, the others are
    gone, and `J5` pin 2 (the harness spare) is tied to `+5V_LED_RING` so the
    pour flows through it to pins 5/6. The pour is now full-width from Q5 to
    `J5`.
  - Check in JLCPCB's placement preview that Q5's pin-1 dot lines up with
    the footprint's pin-1 mark. The footprint geometry and rotation are
    unchanged; only the pad numbers changed.

- **2026-09-27 pre-order fixes** (scripted: `kicad_mainboard/tools/cap_fixes.py`
  then `cap_fixes_pcb.py`; `kicad-cli` DRC afterward: 0 violations,
  0 unconnected, schematic parity unchanged):
  - **`C6` was reversed.** Its + pin was wired to GND and its − pin to
    `+5V_LED_PSU`, in both the schematic and the PCB, so the assembled
    220µF electrolytic would have sat reverse-biased across the 5V input.
    Fixed by rotating the footprint 180° in place (its + pad now lands on
    the existing +5V copper, no re-routing needed).
  - **Added `C8`, 100µF/35V SMD electrolytic on `+24V_MOTOR`** (Lelon
    VZH101M1VTR-0607, LCSC C176683 — Extended part, confirm stock at
    order). The motor rail had no bulk capacitance at all, and `D1` blocks
    regenerated motor energy from returning to the supply. It sits right
    beside the TMC2209's VM pin, with its ground going straight to the
    module's own GND pin (J2-MOTOR pin 2). It's 7.7mm tall, so it clears
    the socketed modules. The gap there is 9.17mm, so it uses a
    project-local copy of `CP_Elec_6.3x7.7` whose courtyard is trimmed to
    0.13mm beyond the pads (`lib.pretty/CP_Elec_6.3x7.7_TrimmedCrtYd`). The
    pads still clear both neighboring socket bodies by more than 0.6mm.
  - **Silkscreen orientation marks** for all three plug-in modules
    ("RJ45/USB END" — since changed to "RJ45 SIDE UP", see 2026-09-28 — / "USB-C END" arrows, "EN"/"VM" at the TMC2209's pin 1
    row), since a module socketed 180° backwards puts power on the wrong
    pins.
  - Gerbers, BOM/CPL, renders, schematic PDF and ERC/DRC reports all
    regenerated.
- **2026-09-27 network firmware update + log** (firmware only; compiles
  clean, **not yet tested on hardware**): `/update` over Ethernet with
  motor lockout, stall timeout and rollback; `/log` viewer backed by a
  12KB ring buffer that every `Log.print` also writes; web UI password
  (Basic auth) on every page once set. First install is one USB flash; see
  [Updating firmware over the network](README.md#updating-firmware-over-the-network).
- **2026-09-27 second pass** (firmware; `tools/add_c9.py` then
  `tools/labels_and_c9_pcb.py`; DRC afterward: 0 violations, 0 unconnected):
  - The first `/scaleEncoder` message after boot or an Ethernet reconnect
    now only seeds the position. It used to be diffed against 0.0, which
    could jump `speedScale` by up to ±0.4 and start the motor backwards.
  - OSC handlers are now registered before `ethConnected` goes true, so
    `oscTask()` can't pump OSC while subscriptions are still being added.
  - The TMC2209's power-up reset flag is cleared at boot. Before, every
    boot logged a false "driver-side reset detected" 5s in.
  - Added `C9`, 0.1µF decoupling at `U1`'s VDD. Nothing on this board's
    `+3V3` rail had a local bypass cap.
  - Silkscreen pin labels outside every hand-wired connector's housing:
    `J-PWR-IN` (a crimp in the wrong cavity could put 24V on the LED ring),
    `J5`, `J3`, a V/S/D/C/G letter per encoder-header pin plus a legend
    line, `J-BTN-TAP` and `J-EXP`. The encoder, tap and expansion headers'
    long reference designators are hidden, replaced by short names
    (MAIN/HUE/…/TAP/EXP) under each header.
- **PCB copper was hand-edited (2026-09-20) to match the encoder pull-up
  fix — now verified with a real KiCad 10.0.6 `kicad-cli` DRC pass, zones
  refilled.** The 8 encoder headers' pull-up pads, and the trace chain
  daisy-chaining them together, were retargeted from net `+5V_LED` to net
  `+3V3` in `Mainboard.kicad_pcb`. The old trace run back to `D2` (which
  fed them from `+5V_LED`) was deleted, so `D2`/`F2` are now a genuinely
  dead-end branch with zero pads left on that net (matches the schematic).
  The now-isolated encoder chain was bridged into the real `+3V3` copper
  mesh with a short F.Cu/B.Cu/F.Cu via-hop from a point above
  `J-ENC-EXP1` that lands exactly on the board's existing I2C-pull-up
  `+3V3` trace, routed around the `TAP_BTN` net's trace it would
  otherwise have crossed. First DRC pass with the straight-line version of
  this bridge genuinely caught two real problems (a track crossing
  `TAP_BTN`, then after moving to B.Cu, crossings/shorts against `GND` and
  `EXP_A0`) — both fixed by rerouting, not by weakening any rule. Final
  result: **`kicad-cli pcb drc --severity-all --schematic-parity
  --refill-zones`: 0 violations, 0 unconnected items** (90 pre-existing
  footprint-value/parity warnings remain, same count as before this fix,
  all pre-dating this change — see `erc_report.rpt`/`drc_report.rpt`).
  `F2`/`D2` are wired but genuinely unused now — **dropped from the BOM**
  (see [`kicad_mainboard/jlc_export/README.md`](mainboard/jlc_export/README.md)),
  though both footprints stay unpopulated on the PCB itself rather than
  being pulled from the layout.
- **Early history of `J1` (before the 2026-09-27 entries).** `J1` used to be one abstract,
  sequentially-numbered `Conn_01x24`, not matching Waveshare's real `H1`/`H2` layout. It grew
  out of fixing a real bug: 14 encoder/button nets (`ENC_CLK/DT/SW`, `HUE_CLK/DT/SW`,
  `SAT_CLK/DT/SW`, `VAL_CLK/DT/SW`, `TAP_BTN`, `DIR_BTN`) dead-ended at their own panel header
  with no wire back to the ESP32 chip, so no encoder or button would have worked. `DIR_BTN` was
  later removed outright (no direction button in the design) and `HUE_SW`/`VAL_SW` moved onto the
  I2C expander. For PCB layout the abstract connector was replaced by the real
  `J1-H1`/`J1-H2` split, and `J2`/`J6` got the same real-footprint treatment. An earlier pass
  through the TMC2209 header also assumed 5 GND pins; the module has 2 (one per row), and
  `DIAG`/`VREF`/`INDEX` are separate solder points, not on either header row.
- **Pre-2026-09-27 design decisions that changed.** The encoder headers' pull-up supply was
  `+5V_LED` until 2026-09-20, which put ~5V on 3.3V-only inputs; all 8 now use `+3V3`, which left
  the general `+5V_LED` branch (`F2` + `D2`) with no load, and it was removed on 2026-09-28.
  `U2` was a 1A TSR 1-2433 until 2026-09-29. The `3V3` rail briefly had a 2A PTC (`F4`) and then
  a 0Ω link before `U2` was wired straight onto `+3V3`.
