Add MicroslopSam TTS, Flextalk voice profiles, effects, and engine controls

CURRENT.md

@@ -1,6 +1,6 @@
 # CURRENT.md — session state
 
-Last updated: 2026-09-14
+Last updated: 2026-09-15
 
 This file tracks the work state. Read it first. Update it before you stop.
 
@@ -12,6 +12,17 @@ mobile draft at 390px.
 
 ## In progress
 
+- Flextalk voice profile, end to end. The profile window is a movable panel that recenters on each
+  open; it carries the character sliders with engine defaults and +/-2 snapping, five named
+  presets, radio-kit effects (high-pass 40-400 Hz, echo 0-100, word mode with a gap in ms), and
+  test lines (Attention, Laboratory, Critical, Health, Radiation, plus two custom lines in
+  `localStorage`) that post the current profile straight to `/api/audio/speak`. Fixed the test
+  buttons reading a stale expression from saved state; they now read the slider like Speak does.
+- Root cause of the profile never reaching the engine (2026-09-15): the wrapper wrote the Flextalk
+  `.reg` with `echo`, and dash's `echo` treats `\2` in `...\Watson\2.0\...` as an escape, which
+  corrupted the key so `wine reg import` was a silent no-op (it still returns 0). The wrapper now
+  uses `printf`, validates that every profile value is numeric, and rejects a `.reg` containing
+  control characters. Documented in TTS-DESIGN.md.
 - The Veskslop agent lives in `agent/` (Go 1.25+, Eino v0.9.19, remote llama.cpp over the
   OpenAI-compatible API, daemon tools over MCP Streamable HTTP). It is disabled by default
   (`VESKSLOP_AGENT_ENABLED=1`), takes tasks on `POST /task`, and logs every tool call. Smoke-tested
@@ -24,6 +35,115 @@ mobile draft at 390px.
   390x844. Next: transcript and TTS box in the Call tab, per the 1c design.
 - Fixed a pre-existing web UI bug: `loadDms` called the nonexistent `renderDirect`, so the DM rail
   only rendered from other paths. It now calls `renderDmRail()`.
+- Fixed the speaking marks: the core turned Discord's numeric `speaking: 0` into `true`, so the
+  equalizer bars never stopped. The bars now fade in on start and decay out on stop (500 ms), with
+  a 2.5 s idle timeout for missed stops; a mute or channel change ends the state at once. Muted
+  members show no bars.
+- Browser audio: **Client audio** now runs the browser unlock in the click (microphone permission,
+  AudioContext created and resumed, output sink applied) before it starts listening. Verified
+  headless: 52 PCM frames played through a running context. The button label no longer wipes the
+  icon markup.
+- Capture routing: nothing moved the client's playback streams into `veskslop_capture`, so the
+  browser monitor streamed digital silence. The audio manager now adopts matching streams
+  (`VESKSLOP_AUDIO_CLIENT_MATCH`, default `vesktop,electron,chromium,discord`) on device setup,
+  monitor start, and every 5 s while listening or capturing.
+- LAN audio verified end to end on 2026-09-14: a tone stream on `veskslop_tts` was moved to
+  `veskslop_capture` (daemon log: `moved stream 81 to veskpop_capture`) and the browser, connected
+  to `https://192.168.0.178:18801/`, played 256+ PCM frames with a peak of 4096 (a stream already
+  on the capture sink peaked at 29274).
+- Speaking bars are always in the layout now: `scaleY(0)` and transparent when idle, they scale up
+  with the fade-in at start, run the equalizer animation while speaking, and stay on until an
+  explicit stop, a mute, or a channel change. Live data showed repeated `speaking: true` pulses and
+  no stop events inside an 8 s window, so the idle decay is removed: bars stick rather than
+  disappear. The fade runs the equalizer inside a shrinking container, so the amplitude drops
+  smoothly instead of the block collapsing at full height. The core treats a missing `speaking`
+  flag as a start pulse and accepts `user_id`. Core-side changes need a rebuild plus a client
+  restart.
+- Added `poc/scripts/audio-doctor.sh` (`pnpm audio:doctor`): default sink, sink inputs, audio
+  status with adoption counters, a 5 s `volumedetect` on `veskslop_capture.monitor`, and a live
+  event peek. Run it in a real call to trace a silent browser monitor.
+- Live doctor output (2026-09-14): the `vesktop`/`electron` streams sit on `veskslop_capture`
+  (sink 40) already, `adopted: 0`, and `VOICE_SPEAKING` events arrive. Client audio routing is
+  therefore in place. Bot TTS works. Capture was off (`capturing: false`), so the transcript was
+  empty; Vosk transcribes correctly when invoked with the model env.
+- Live call audio works now (the user confirmed TTS is audible and the silent call was partly a
+  per-user volume set to zero for some members, plus another issue). The `volume` read returns
+  `null` because this client build exposes no `UserVolumeStore`; the setter reports
+  `NOT_SUPPORTED`. Volume is not a usable control here.
+- External TTS engine support (external project, first target is a command wrapper): the contract
+  lives in `TTS-DESIGN.md`. Veskslop now has nullable `tts.speed` (0.5-2.0) and `tts.pitch`
+  (-12..+12) preferences, per-line overrides on `/api/audio/speak` and MCP `speak`, `{speed}`/
+  `{pitch}` placeholders, `VESKSLOP_TTS_SPEED`/`_PITCH` exports, and speed/pitch sliders in the TTS
+  panel. A custom command advertises its voices through `VESKSLOP_TTS_VOICES`; the espeak list is
+  hidden while a custom command is set. Verified with a wrapper that logged argv and env:
+  preferences and overrides both arrive. Reading the engine's `GET /voices` for the HTTP provider
+  remains future work.
+- The MicroslopSam engine project is now part of the tree in `microslopsam/` (copied from the
+  separate project; its `build/` output and voice data stay ignored). The wrapper was renamed
+  `setup/vekslop-tts.sh` -> `setup/veskslop-tts.sh` and every reference and configuration export
+  was fixed. The engine also corrected the Flextalk pitch finding (REG_DWORD bug), so Keith Bell
+  and Kerry Watson now support pitch and carry no `no-pitch` flag.
+- Implemented the Flextalk Expression profile parameter end to end: `VESKSLOP_TTS_VOICES` marks the
+  two voices as `expression=59` / `expression=100`; the wrapper writes the registry value before
+  synthesis; Veskslop has `tts.expression` (0-100 nullable), the `{expression}` placeholder,
+  `VESKSLOP_TTS_EXPRESSION`, an HTTP/MCP field, and an Expression slider that appears only for
+  those voices (default shown until moved, null leaves the registry untouched). Verified: 0/59/100
+  produce distinct audio, the registry reaches 0x14 for 20, and the UI shows 25/100/default states.
+- Fixed the null-value coercion in the preference normalizers (`Number(null) === 0` turned a
+  cleared speed/expression into 0.5/0); unset now stays null.
+- Renderer-down handling polished: the bridge logs one warning per outage instead of a stack per
+  polled command, and the `navigator.languages` request catches the new timeout rejection instead
+  of raising an unhandled promise warning. (The `exitCode: 9` renderer kills in the live log were
+  client restarts; the daemon's retry handled the startup race as designed.)
+- The TTS panel has speed presets (0.5x, 0.75x, 1x, 1.25x, 1.5x) under the fine-adjust slider, and
+  the pitch control is hidden entirely for voices that report `supportsPitch: false`. (Flextalk
+  voices no longer report that; see the pitch correction above.) Verified: preset clicks update the
+  slider/label and save.
+- TTS playback has a lead-in of silence before each line (`VESKSLOP_TTS_LEAD_MS`, default 150 ms,
+  max 2000, added with `adelay` in `ffplay`; status reports `ttsLeadMs`) so Discord's voice
+  pipeline is settled before the first syllable. Verified on the sink monitor: first sound at
+  0.02 s with 0 ms and 0.62 s with 600 ms.
+- Speech settings have a **Start engine** button that warms the selected engine
+  (`POST /api/audio/tts/warm`, MCP `start_tts_engine`): it runs one throwaway synthesis so the
+  command engine's Xvfb/Wine/host are up. Verified with MicroslopSam: cold 1732 ms, warm 505 ms,
+  and the UI shows "MicroslopSam ready in 835 ms". The warm text is a full sentence because their
+  two-pass `loudnorm` fails with `inf` on very short input.
+- Startup and crash robustness: `sendRendererCommand` now rejects immediately when the client
+  renderer is destroyed or crashed, and times out after 15 s, so a dead frame no longer logs
+  Electron's "Render frame was disposed" stack once per poll. Device auto-connect retries up to
+  10 times every 2.5 s, because the bridge socket is ready before the renderer's modules load
+  (the live error was `Reflect.get called on non-object` from a lazy lookup at startup). Verified
+  with the mock: attempt 1 failed, attempt 2 set the devices.
+- TTS engines are selectable: the Speech settings and the TTS panel list espeak-ng, the command
+  engine (`VESKSLOP_TTS_NAME`, default "custom engine"), and the HTTP API; unavailable entries are
+  disabled and Auto resolves to the command engine first. Selecting an engine switches the voice
+  list (`/api/audio/tts/voices?engine=`) and hides the language row for single-language engines.
+  The TTS panel now has a voice select next to the speed/pitch sliders. Fixed a slider bug where a
+  null speed preference (`Number(null) === 0`) clamped the slider to 0.5 and then persisted it.
+- TTS controls: the TTS panel has a working **duck on voice** toggle (stored in preferences,
+  reported as `duckEnabled`). Ducking was verified live with the mock: the `veskslop_tts` sink
+  drops to 30% while TTS plays and someone speaks, and returns to 100% after. The Transcript tab
+  in the console has **Start** and **Stop** buttons; the status line shows `capturing`/`capture
+  off`. The espeak-ng voice list now includes the `m1`–`m7`, `f1`–`f5`, `croak`, and `whisper`
+  variants (`<voice>+<variant>`), so each language has about 15 voices instead of one.
+- Speaking state is now authoritative: the daemon polls the core's `voice.speaking` action
+  (`ChannelRTCStore.getSpeakingParticipants`) every second and synthesizes the missing stop/start
+  events; dispatcher `VOICE_SPEAKING` hints feed the same set so stale users clear. This fixes the
+  bars sticking for everyone and the transcript attribution, which credited the first member
+  because speaker intervals never closed. Verified with the mock: a `speaking: true` event cleared
+  within one poll.
+- The client streams sit uncorked, unmuted, and at 100% on `veskslop_capture`, yet the monitor is
+  silent. Gateway `VOICE_SPEAKING` events do not prove inbound RTP, and the TTS path only proves
+  outbound, so the remaining suspect is the client's inbound voice state (stuck deafen or a broken
+  voice connection), not the sink routing. `/health` now returns `selfId` so the bot's own member
+  flags can be picked out of the voice member list; the doctor prints it.
+- Client output is still silent at the sink (`-91 dB`) although the streams are uncorked, unmuted,
+  and at 100%, and a headless Chromium tone into the same sink measured `-7.4 dB`. Added per-user
+  volume reading and writing to find Discord-side silencing: `user.volume` / `user.setVolume` in
+  the core, `/api/voice/volume` (GET and POST), MCP `get_user_volume` / `set_user_volume`, and
+  `volume` in every voice member. A volume of 0 (or an output volume of 0) silences a user even
+  when Pulse shows a healthy stream. `voice.join` also clears persisted self-deafen now.
+  The doctor prints output and per-user volumes.
 - Voice audio is implemented: daemon-managed virtual devices (`veskslop_tts`, `veskslop_mic`,
   `veskslop_capture`), TTS providers (local command + HTTP API), queued playback through `ffplay`,
   ducking on `VOICE_SPEAKING`, and WAV capture. The daemon starts `pulseaudio` when `pactl` cannot