Expose vexa's STT Whisper chunking to .env and increase the window

stt-window-tuning.patch

diff --git a/core/meetings/modules/mixed-pipeline/src/chunked-transcriber.ts b/core/meetings/modules/mixed-pipeline/src/chunked-transcriber.ts
index f37484f5..402c6784 100644
--- a/core/meetings/modules/mixed-pipeline/src/chunked-transcriber.ts
+++ b/core/meetings/modules/mixed-pipeline/src/chunked-transcriber.ts
@@ -51,6 +51,13 @@ import type { TranscriptionResult } from '@vexa/transcribe-whisper';
 import { localAgreement } from '@vexa/transcribe-buffer';
 
 const SAMPLE_RATE = 16000;
+/** Operator-set positive-number knob. A missing, empty, or malformed value falls back to the
+ *  built-in default: a NaN threshold would compare false everywhere and silently wedge the
+ *  submit path, so a fat-fingered env var must never be load-bearing. */
+const envMs = (raw: string | undefined, fallback: number): number => {
+  const n = Number(raw);
+  return raw !== undefined && raw.trim() !== '' && Number.isFinite(n) && n > 0 ? n : fallback;
+};
 /** Near-silent spans are dropped before Whisper (desktop's DROP_RMS). */
 const DROP_RMS = 0.006;
 /** Ring capacity — must hold a full unconfirmed window plus segmenter lag. */
@@ -88,13 +95,16 @@ const SILENCE_CLOSE_CONTEXT_MS = Number((typeof process !== 'undefined' && proce
  *  sentence pauses, below an awkward wait. */
 const CONFIRM_TTL_MS = 2500;
 /** Don't bother Whisper with unconfirmed windows shorter than this unless
- *  the turn is closing. */
-const MIN_SUBMIT_MS = 800;
+ *  the turn is closing. Raising it buys the decoder more acoustic context per
+ *  call (fewer, longer windows) at the cost of a later first pending — worth it
+ *  on languages where short windows mis-decode. Tunable via BOT_MIXED_MIN_SUBMIT_MS. */
+const MIN_SUBMIT_MS = envMs(typeof process !== 'undefined' ? process.env?.BOT_MIXED_MIN_SUBMIT_MS : undefined, 800);
 /** Time-based resubmission cadence for the OPEN turn (the bot's
  *  submitInterval): pending refreshes and LocalAgreement stability build at
  *  this pace instead of waiting for the next boundary (which can be 10s away
- *  inside a monologue). Pending ≈ tick + RTT; confirm ≈ 2 ticks. */
-const SUBMIT_TICK_MS = 2000;
+ *  inside a monologue). Pending ≈ tick + RTT; confirm ≈ 2 ticks — so raising it
+ *  trades confirm latency for context. Tunable via BOT_MIXED_SUBMIT_TICK_MS. */
+const SUBMIT_TICK_MS = envMs(typeof process !== 'undefined' ? process.env?.BOT_MIXED_SUBMIT_TICK_MS : undefined, 2000);
 /** A short isolated active-speaker UI switch (Zoom/Teams) right after a different
  *  published speaker is held provisional rather than stamped — without acoustic
  *  evidence a brief tile flip is more likely a stale/echoed hint than a real,
diff --git a/core/runtime/src/runtime_kernel/profiles.py b/core/runtime/src/runtime_kernel/profiles.py
index 5df9cea0..479e3a66 100644
--- a/core/runtime/src/runtime_kernel/profiles.py
+++ b/core/runtime/src/runtime_kernel/profiles.py
@@ -105,6 +105,8 @@ def default_registry() -> ProfileRegistry:
             "BOT_SPEAKER_CONFIRM_THRESHOLD",
             "BOT_SPEAKER_MAX_BUFFER_SEC",
             "BOT_SPEAKER_IDLE_TIMEOUT_SEC",
+            "BOT_MIXED_MIN_SUBMIT_MS",
+            "BOT_MIXED_SUBMIT_TICK_MS",
         )
         if os.environ.get(key, "").strip()
     }
diff --git a/deploy/compose/.env.example b/deploy/compose/.env.example
index 2f86c1ac..13019e9e 100644
--- a/deploy/compose/.env.example
+++ b/deploy/compose/.env.example
@@ -122,6 +122,11 @@ BOT_SPEAKER_SUBMIT_INTERVAL_SEC=
 BOT_SPEAKER_CONFIRM_THRESHOLD=
 BOT_SPEAKER_MAX_BUFFER_SEC=
 BOT_SPEAKER_IDLE_TIMEOUT_SEC=
+# Mixed lane (Zoom/Teams/Jitsi) STT windowing. Defaults 800 / 2000 ms. Raise both to give the
+# decoder more acoustic context per Whisper call — fewer, longer windows decode short words more
+# reliably (notably in German), at the cost of a later pending (≈ tick) and confirm (≈ 2 ticks).
+BOT_MIXED_MIN_SUBMIT_MS=
+BOT_MIXED_SUBMIT_TICK_MS=
 # Self-hosted Jitsi hostnames (comma-separated, e.g. calls.example.org) recognized when parsing
 # pasted meeting links AND calendar (ICS) links. meet.jit.si and hosts naming "jitsi" are always
 # recognized; hosts with a "meet" label (meet.example.org, eu.meet.example.org) are recognized in
diff --git a/deploy/compose/docker-compose.yml b/deploy/compose/docker-compose.yml
index 50f4a763..49b0cf3b 100644
--- a/deploy/compose/docker-compose.yml
+++ b/deploy/compose/docker-compose.yml
@@ -148,6 +148,8 @@ services:
       - BOT_SPEAKER_CONFIRM_THRESHOLD=${BOT_SPEAKER_CONFIRM_THRESHOLD:-}
       - BOT_SPEAKER_MAX_BUFFER_SEC=${BOT_SPEAKER_MAX_BUFFER_SEC:-}
       - BOT_SPEAKER_IDLE_TIMEOUT_SEC=${BOT_SPEAKER_IDLE_TIMEOUT_SEC:-}
+      - BOT_MIXED_MIN_SUBMIT_MS=${BOT_MIXED_MIN_SUBMIT_MS:-}
+      - BOT_MIXED_SUBMIT_TICK_MS=${BOT_MIXED_SUBMIT_TICK_MS:-}
       - VEXA_AGENT_SRC_MOUNT=${VEXA_AGENT_SRC_MOUNT:-}
       - REDIS_URL=redis://redis:6379/0
       # The Runtime brokers model credentials into spawned agents. Subscription credentials may be