…humansai#1206)

Improves voice interaction reliability and observability across the TTS,
STT, and mascot layers:

audioPlayer.ts
- Add `PlaybackOptions.maxDurationMs` — auto-stops runaway TTS after a
  configurable deadline (default not set; mascot passes TTS_MAX_PLAYBACK_MS
  = 5 min) so the mascot can't get stuck in a permanent `speaking` pose
- Add `debug('human:audio-player')` lifecycle logging: playback start,
  metadata-ready duration, natural end, auto-stop with reason

useHumanMascot.ts
- Export `TTS_MAX_PLAYBACK_MS` (5 min) and pass it to `playBase64Audio`
  via the new `PlaybackOptions` interface
- Improve the `listening` interrupt log: distinguish "user started recording
  while TTS was playing (interrupted)" from "mic activated, no TTS to cancel"
  with a stable `tts-in-flight=true/false` field for grepping

MicComposer.tsx
- Add `retryCount` state: tracks how many STT retry attempts have been made
  for the current recording session
- Export `STT_MAX_RETRIES` so tests and callers can reference the ceiling
- Thread `onRetry` callback through `transcribeWithRetry` — each transient
  backoff increments `retryCount` so the button label can change
- Update label: `transcribing + retryCount > 0` now renders
  `mic.retryingTranscription` ("Retrying… (1 of 2)") instead of the
  generic "Transcribing…" spinner text
- Add per-session monotonic ID and structured `[session:N]` log prefix
  so recording start/stop/retry/error events for the same session are
  trivially correlated in debug output

i18n
- Add `mic.retryingTranscription` key ("Retrying... ({attempt} of {max})")
  to en.ts and all 14 locale *-2.ts chunk files (English placeholder)

Tests (13 new across three suites, 109 total)
- audioPlayer: `maxDurationMs` auto-stops and rejects `ended`; timer cleared
  on natural end before deadline
- MicComposer: `STT_MAX_RETRIES` is a positive integer (contract test)
- useHumanMascot: listening override cancels ack timer and transitions to
  listening face without timer firing late; `TTS_MAX_PLAYBACK_MS` sanity

Closes tinyhumansai#1206

1. audioPlayer.ts: replace unsupported %.2f precision specifier with
   audio.duration.toFixed(2) passed as %s — the debug package only
   supports %O/%o/%s/%d/%j/%%.

2. MicComposer.tsx: fix cumulative retry counter double-counting on the
   WAV fallback path. Before the fix, handleRetry mutated cumulativeRetries
   and the WAV callback added cumulativeRetries (already updated) to the
   per-path attempt number, producing the sequence 1,2,3,5 instead of
   1,2,3. Fix: snapshot nativeRetries before the WAV path so the closure
   captures a stable value. Also corrected {max} in the retry label from
   STT_MAX_RETRIES (2) to STT_MAX_RETRIES * 2 (4) — the actual cumulative
   ceiling across both codec paths.

3. MicComposer.test.tsx: add test that verifies the retry sequence is
   [1, 2, 3] (not [1, 2, 3, 5]) when both native and WAV paths retry, and
   that 5 total transcribe calls are made.

…ession

Replace hardcoded literal pushes (which were independent of the component's
computed state) with a DOM read inside the final WAV success mock call.

By the time WAV attempt 2 (call 6) executes, React has flushed the
setRetryCount update from WAV retry-1's callback through the 500ms
fake-timer backoff (act-wrapped). Reading the rendered <span> text there
gives the component's actual retryCount value:
  fixed:  "Retrying... (4 of 4)"  (nativeRetries snapshot = 2, 2+2=4)
  buggy:  "Retrying... (5 of 4)"  (mutable cumulativeRetries=3, 3+2=5)

Verified: reverting the nativeRetries snapshot fix causes this test to fail.

…olithic locale files

Upstream (tinyhumansai#3004) deleted all i18n chunk files and merged them into
per-locale monolithic files (ar.ts, bn.ts, …). The mic.retryingTranscription
key added by this PR was in the now-deleted *-2.ts chunks; moved it to the
corresponding line (after mic.transcribing) in each of the 13 locale files.
en.ts already contained the key from the auto-merge.