fix(bot): pass the measured 800 ms CSRC inactivity window from the composition root

core/meetings/modules/mixed-capture-core/src/csrc-poll.test.ts

@@ -85,7 +85,11 @@ const receiver = (sources: () => ContributingSourceLike[]) => ({
   check('past inactiveMs the deactivation is SYNTHESIZED (the transport never sends this edge)',
     out.length === 2 && out[1].active === false && out[1].csrc === 7 && out[1].tMs === t,
     JSON.stringify(out));
-  check('the default inactivity window is 400ms', CSRC_INACTIVE_MS === 400);
+  // The module's own default spans a PACKET gap (jitter, DTX), not a speech PAUSE. It is only a
+  // default: the measured production window is longer and is passed in by the composition root,
+  // so what this file owns is the SEAM, not the number production runs on.
+  check('the built-in window is the packet-gap default, and nothing more than a default',
+    CSRC_INACTIVE_MS === 400, String(CSRC_INACTIVE_MS));
 
   // Speaking again is a NEW activation — turns are edges, not a level.
   speaking = true; lastSpoke = t; poll.poll();
@@ -100,6 +104,33 @@ const receiver = (sources: () => ContributingSourceLike[]) => ({
   check('a poll after destroy() emits nothing', out.length === after);
 }
 
+// ── the inactivity window is a PARAMETER, not a constant ─────────────────────────────────
+// A caller with a MEASURED window must be able to hand it over instead of forking this file —
+// the seam, not a copied constant, is the contract.
+{
+  let t = 1_900_000_000_000;
+  const out: CsrcTransition[] = [];
+  let speaking = true;
+  let lastSpoke = t;
+  const poll = createCsrcPoll({
+    onTransition: (x) => out.push(x),
+    now: () => t,
+    timeOrigin: () => 0,
+    inactiveMs: 800,
+    receivers: () => [receiver(() => [{ source: 9, timestamp: speaking ? t : lastSpoke, audioLevel: 0.4 }])],
+  });
+  poll.poll();
+  check('override: the source opens a turn', out.length === 1 && out[0].active === true, JSON.stringify(out));
+  speaking = false;
+  t += 500; poll.poll();
+  check('override: 500ms of silence — past the 400ms default, inside the caller\'s 800ms — holds the turn OPEN',
+    out.length === 1, JSON.stringify(out));
+  t += 400; poll.poll();   // 900ms > 800ms
+  check('override: past the caller\'s window the deactivation is synthesized',
+    out.length === 2 && out[1].active === false, JSON.stringify(out));
+  poll.destroy();
+}
+
 // ── two concurrent sources ──────────────────────────────────────────────────────────────────────
 {
   let t = 1_900_000_000_000;

core/meetings/services/bot/src/capture-bridge.ts

@@ -381,6 +381,45 @@ const TEAMS_ENABLE_CAPTIONS = process.env.VEXA_TEAMS_ENABLE_CAPTIONS === '1';
 /** How many times to try the menu path before giving up (each attempt is ~3 s of UI waits). */
 const TEAMS_ENABLE_CAPTIONS_ATTEMPTS = Math.max(1, Number(process.env.VEXA_TEAMS_ENABLE_CAPTIONS_ATTEMPTS || 3));
 
+/** The sensor's poll cadence (`CSRC_POLL_MS`). Not imported: this file must not pull the page-side
+ *  bundle into the Node build. Kept as the floor only, so a drift there is a slacker floor here,
+ *  never a wrong window. */
+const CSRC_MIN_INACTIVE_MS = 100;
+/** No plausible window is 10 s: the measurement below shows 1600 ms already merges distinct
+ *  turns, so a huge override is a typo wearing a number — the NaN failure mode again, silent. */
+const CSRC_MAX_INACTIVE_MS = 10_000;
+/**
+ * The measured window — how long the transport sensor holds a source active after its last
+ * observed contribution. One name, so the fallback, the message and the default cannot drift.
+ *
+ * The sensor's own default is 400 ms (`CSRC_INACTIVE_MS` in @vexa/mixed-capture-core) — a value
+ * sized to span a packet gap. A speech PAUSE is longer: entry-timestamp staleness during natural
+ * pauses measures p50 550 ms, so at 400 ms the MEDIAN pause trips a synthesized deactivation and
+ * one speech phase fragments into splinters. Against the current channelizer (5 × 30 min seeded
+ * tapes): 400 ms yields 1.67 lane activations per real turn (67 % over-splitting), 800 ms yields
+ * 1.13 — the knee — for 1.8 pp more contamination; 1600 ms starts merging distinct turns. So the
+ * production window is 800 ms.
+ *
+ * The number lives HERE, at the composition root; the sensor owns the `inactiveMs` seam and its
+ * packet-gap default, nothing more. Overridable via VEXA_CSRC_INACTIVE_MS where the bot process
+ * env is operator-controlled: the runtime kernel forwards it to spawned bots (profiles.py's
+ * tuning allowlist), and the lite/process backend inherits the host env directly.
+ */
+const CSRC_DEFAULT_INACTIVE_MS = 800;
+export function resolveCsrcInactiveMs(raw: string | undefined, warn: (m: string) => void = () => { /* silent */ }): number {
+  // `undefined` is "no override", which is the normal case and silent. An override that was SET —
+  // including an empty or blank one, which is what a half-finished deploy template looks like — is
+  // an intent that could not be honoured, and gets said out loud.
+  if (raw === undefined) return CSRC_DEFAULT_INACTIVE_MS;
+  const n = Number(raw);
+  if (!Number.isFinite(n) || n < CSRC_MIN_INACTIVE_MS || n > CSRC_MAX_INACTIVE_MS) {
+    warn(`[bot] VEXA_CSRC_INACTIVE_MS=${JSON.stringify(raw)} is not a usable window (finite, ${CSRC_MIN_INACTIVE_MS}..${CSRC_MAX_INACTIVE_MS}ms) — using the measured ${CSRC_DEFAULT_INACTIVE_MS}ms`);
+    return CSRC_DEFAULT_INACTIVE_MS;
+  }
+  return n;
+}
+const CSRC_RESOLVED_INACTIVE_MS = resolveCsrcInactiveMs(process.env.VEXA_CSRC_INACTIVE_MS, (m) => console.warn(m));
+
 /** Outcome of one enable attempt. `already-on` and `clicked` are successes; `failed` carries WHY,
  *  because "captions never appeared" has two very different causes — the menu path changed, or the
  *  tenant blocks captions — and only the reason distinguishes them on the first live run. */
@@ -796,7 +835,7 @@ export async function startCaptureBridge(
   // ── Start the page-side capture (VexaBrowserUtils preferred; production inline fallback). ──
   // The body of this callback runs IN THE BROWSER (Playwright serializes it); DOM globals are
   // reached via globalThis (this file type-checks against the Node lib — no DOM types here).
-  await page.evaluate(async ({ isMixed, isPerTrack, isJitsi, isTeams, isZoom, botName, mainAudioGraceMs, mainAudioSilenceMs, mainAudioEnergyRms }) => {
+  await page.evaluate(async ({ isMixed, isPerTrack, isJitsi, isTeams, isZoom, botName, mainAudioGraceMs, mainAudioSilenceMs, mainAudioEnergyRms, csrcInactiveMs }) => {
     const w = (globalThis as any) as Record<string, any>;
     if (isMixed) {
       // Zoom/Teams/Jitsi ride the WebRTC hook (installRemoteAudioHook, installed pre-nav), which mirrors
@@ -1159,6 +1198,9 @@ export async function startCaptureBridge(
               w.logBot?.('[Csrc] observation ' + JSON.stringify(o));
               w.__vexaObservation?.('csrc', o, Date.now());
             },
+            // The MEASURED pause window, passed in rather than defaulted: the sensor's own 400 ms
+            // is shorter than the median natural speech pause (p50 550 ms) and fragments a turn.
+            inactiveMs: csrcInactiveMs,
             log: (m: string) => w.logBot?.('[Csrc] ' + m),
           });
         } catch (e: any) {
@@ -1287,7 +1329,9 @@ export async function startCaptureBridge(
       mainAudioGraceMs: Number(process.env.VEXA_TEAMS_MAIN_AUDIO_GRACE_MS || 15000),
       // How long a PICKED mix may stay wholly silent before the lane abandons it for every track.
       mainAudioSilenceMs: Number(process.env.VEXA_TEAMS_MAIN_AUDIO_SILENCE_MS || 20000),
-      mainAudioEnergyRms: Number(process.env.VEXA_TEAMS_MAIN_AUDIO_ENERGY_RMS || 0.006) }).catch((e) => {
+      mainAudioEnergyRms: Number(process.env.VEXA_TEAMS_MAIN_AUDIO_ENERGY_RMS || 0.006),
+      // The measured inactivity window — a parameter the sensor accepts, never a fork of it.
+      csrcInactiveMs: CSRC_RESOLVED_INACTIVE_MS }).catch((e) => {
     console.error(`[bot] capture bridge: page-side start failed: ${String(e)}`); // L4: surfaces only on the VM
   });
 

core/meetings/services/bot/src/csrc-wiring.test.ts

@@ -24,13 +24,19 @@
 import { readFileSync, existsSync } from 'node:fs';
 import { fileURLToPath } from 'node:url';
 import path from 'node:path';
-import {
-  startCaptureBridge, makeCsrcSink, makeObservationSink,
-  type CsrcRecord, type CsrcCapableSink, type ObservationRecord, type ObservationCapableSink,
+import type {
+  CsrcRecord, CsrcCapableSink, ObservationRecord, ObservationCapableSink,
 } from './capture-bridge.js';
 import type { Invocation } from './config.js';
 import type { BotPipeline } from './pipeline.js';
 
+// HERMETIC: capture-bridge resolves VEXA_CSRC_INACTIVE_MS at module load, so an ambient override
+// in the developer's shell would fail the 800ms assertions below. Scrub BEFORE the module loads —
+// which is why this import is dynamic (a static one hoists above the delete).
+delete process.env.VEXA_CSRC_INACTIVE_MS;
+const { startCaptureBridge, makeCsrcSink, makeObservationSink, resolveCsrcInactiveMs } =
+  await import('./capture-bridge.js');
+
 let failed = 0;
 const check = (name: string, cond: boolean, detail?: string) => {
   console.log(`  ${cond ? '✅' : '❌'} ${name}${cond || !detail ? '' : ` — ${detail}`}`);
@@ -87,10 +93,57 @@ const check = (name: string, cond: boolean, detail?: string) => {
     warnings.some((w) => w.includes('observation-clock-skew')) && stored[2].t >= t, warnings.join(' | '));
 }
 
+// ── the window is resolved from env, and garbage never reaches the sensor ───────────────────────
+// The sensor resolves its option with `??`, which passes NaN through, and a NaN window makes BOTH
+// of its comparisons false: every source is re-deactivated on every poll — the fragmentation this
+// value exists to end, worse and silent. So an unusable override falls back rather than propagating.
+{
+  check('window: no override ⇒ the measured 800ms', resolveCsrcInactiveMs(undefined) === 800,
+    String(resolveCsrcInactiveMs(undefined)));
+  check('window: a usable override wins', resolveCsrcInactiveMs('1200') === 1200,
+    String(resolveCsrcInactiveMs('1200')));
+  check('window: garbage falls back instead of becoming NaN',
+    resolveCsrcInactiveMs('abc') === 800, String(resolveCsrcInactiveMs('abc')));
+  check('window: zero, negative and empty fall back too — none of them is a window',
+    resolveCsrcInactiveMs('0') === 800 && resolveCsrcInactiveMs('-5') === 800 && resolveCsrcInactiveMs('') === 800,
+    `${resolveCsrcInactiveMs('0')}/${resolveCsrcInactiveMs('-5')}/${resolveCsrcInactiveMs('')}`);
+  // A window shorter than one poll is stale before the next tick — the NaN failure mode wearing an
+  // ordinary number. It is rejected at the floor, not accepted because it happens to be positive.
+  check('window: shorter than one 100ms poll is rejected, one poll exactly is kept',
+    resolveCsrcInactiveMs('1') === 800 && resolveCsrcInactiveMs('99') === 800 && resolveCsrcInactiveMs('100') === 100,
+    `${resolveCsrcInactiveMs('1')}/${resolveCsrcInactiveMs('99')}/${resolveCsrcInactiveMs('100')}`);
+  // The ceiling is the floor's mirror: a fat-fingered 600000 would hold every source active for
+  // ten minutes and synthesize essentially no deactivations — the NaN failure mode wearing an
+  // ordinary number, equally silent. 1600ms already merges turns; 10s is generous headroom.
+  check('window: a huge override is rejected at the ceiling, ten seconds exactly is kept',
+    resolveCsrcInactiveMs('600000') === 800 && resolveCsrcInactiveMs('10001') === 800
+    && resolveCsrcInactiveMs('10000') === 10000,
+    `${resolveCsrcInactiveMs('600000')}/${resolveCsrcInactiveMs('10001')}/${resolveCsrcInactiveMs('10000')}`);
+  {
+    const warnings: string[] = [];
+    const used = resolveCsrcInactiveMs('abc', (m) => warnings.push(m));
+    check('window: a rejected override is SAID OUT LOUD, never silently ignored',
+      used === 800 && warnings.length === 1 && warnings[0]!.includes('VEXA_CSRC_INACTIVE_MS="abc"'),
+      JSON.stringify(warnings));
+    const quiet: string[] = [];
+    resolveCsrcInactiveMs(undefined, (m) => quiet.push(m));
+    resolveCsrcInactiveMs('1200', (m) => quiet.push(m));
+    check('window: no override and a good override are both silent — only a REJECTION warns',
+      quiet.length === 0, JSON.stringify(quiet));
+    // An empty override is what a half-rendered deploy template looks like. Treating it as "not
+    // set" is how a knob gets ignored in silence — the exact failure the warning exists to prevent.
+    const blank: string[] = [];
+    check('window: an override SET to empty or blank warns too — it is an intent, not an absence',
+      resolveCsrcInactiveMs('', (m) => blank.push(m)) === 800
+      && resolveCsrcInactiveMs('   ', (m) => blank.push(m)) === 800
+      && blank.length === 2, JSON.stringify(blank));
+  }
+}
+
 // ── The real bundle (built by build-browser-utils.mjs — turbo test depends on build) ─────────────
 const BUNDLE = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '..', 'dist', 'browser-utils.global.js');
 if (!existsSync(BUNDLE)) {
-  console.error(`❌ missing ${BUNDLE} — build the capture bundle first (pnpm --filter @vexa/bot build).`);
+  console.error(`❌ missing ${BUNDLE} — build the capture bricks first (pnpm --filter @vexa/mixed-capture-core --filter @vexa/bot build).`);
   process.exit(1);
 }
 
@@ -138,6 +191,19 @@ const utils = g.VexaBrowserUtils as Record<string, unknown> | undefined;
 check('bundle: window.VexaBrowserUtils.createCsrcPoll is exported (RED at base — brick not bundled)',
   typeof utils?.createCsrcPoll === 'function', `keys: ${Object.keys(utils ?? {}).join(',')}`);
 
+// Capture the options the REAL bridge hands the REAL sensor factory. The window the production
+// path runs on is not observable from a transition alone (only from WHEN one is absent), and the
+// number itself is owned by the composition root, not by the sensor — so it is asserted here, on
+// the object that actually crosses. RED when the bridge passes no `inactiveMs` at all.
+type CsrcPollOpts = { inactiveMs?: number; pollMs?: number; now?: () => number; timeOrigin?: () => number;
+  receivers?: () => unknown[]; onTransition: (t: { csrc: number; active: boolean; tMs: number }) => void };
+const realCreateCsrcPoll = utils!.createCsrcPoll as (o: CsrcPollOpts) => unknown;
+let productionCsrcOpts: CsrcPollOpts | undefined;
+(utils as Record<string, unknown>).createCsrcPoll = (o: CsrcPollOpts): unknown => {
+  productionCsrcOpts = o;
+  return realCreateCsrcPoll(o);
+};
+
 // ── Fake Playwright Page + the Node seams ───────────────────────────────────────────────────────
 const page = {
   async exposeFunction(name: string, fn: unknown): Promise<void> { g[name] = fn; },
@@ -213,8 +279,41 @@ check('the spine and the stored sidecar agree on WHEN each edge happened',
 check('isolation: NO transition reached pipeline.recordHint — a csrc is an id, never a name',
   hints.length === 0, JSON.stringify(hints));
 
+// ── the MEASURED inactivity window reaches the sensor ────────────────────────────────────────────
+// 400 ms — the sensor's own default — is shorter than the median natural speech pause (measured
+// p50 550 ms), so a turn fragments into 1.67 lane activations. The composition root passes the
+// measured 800 ms instead. Asserted twice: the value that crossed, and what that value DOES.
+check('the bridge passes an explicit inactivity window to the sensor (RED at base: undefined)',
+  productionCsrcOpts?.inactiveMs === 800, `inactiveMs=${String(productionCsrcOpts?.inactiveMs)}`);
+{
+  // Re-drive the sensor over the PRODUCTION options object — same `inactiveMs`, with only the
+  // clock and the receivers replaced, so the window is proven by behaviour and not by a number.
+  let t = 1_900_000_000_000;
+  const edges: Array<{ active: boolean }> = [];
+  let speaking = true;
+  let lastSpoke = t;
+  const poll = realCreateCsrcPoll({
+    ...productionCsrcOpts!,
+    onTransition: (x) => edges.push({ active: x.active }),
+    now: () => t,
+    timeOrigin: () => 0,
+    receivers: () => [{ track: { kind: 'audio' },
+      getContributingSources: () => [{ source: 5, timestamp: speaking ? t : lastSpoke, audioLevel: 0.4 }] }],
+  }) as { poll(): void; destroy(): void };
+  poll.poll();
+  speaking = false;
+  t += 500; poll.poll();
+  check('a 500ms pause — past the sensor default, inside the measured window — does NOT close the turn',
+    edges.length === 1 && edges[0].active === true, JSON.stringify(edges));
+  t += 400; poll.poll();
+  check('past the measured window the deactivation is synthesized',
+    edges.length === 2 && edges[1].active === false, JSON.stringify(edges));
+  poll.destroy();
+}
+
 (g as any).setInterval = realSetInterval;
 (g as any).clearInterval = realClearInterval;
+(utils as Record<string, unknown>).createCsrcPoll = realCreateCsrcPoll;
 g.document = savedDocument;
 
 if (failed) { console.error(`\n❌ csrc-wiring: ${failed} checks FAILED.`); process.exit(1); }

core/runtime/src/runtime_kernel/profiles.py

@@ -142,6 +142,10 @@ def default_registry() -> ProfileRegistry:
             "BOT_SPEAKER_CONFIRM_THRESHOLD",
             "BOT_SPEAKER_MAX_BUFFER_SEC",
             "BOT_SPEAKER_IDLE_TIMEOUT_SEC",
+            # CSRC transport-sensor inactivity window (capture-bridge resolves and guards it) —
+            # without this forwarding, an operator's override on the runtime container would
+            # silently never reach the spawned bot process.
+            "VEXA_CSRC_INACTIVE_MS",
         )
         if os.environ.get(key, "").strip()
     }

core/runtime/tests/test_profiles.py

@@ -61,12 +61,16 @@ def test_meeting_bot_forwards_speaker_stream_tuning(monkeypatch):
     monkeypatch.setenv("BOT_ALONE_SILENCE_WINDOW_MS", "60000")
     monkeypatch.setenv("BOT_SPEAKER_MIN_AUDIO_SEC", "1")
     monkeypatch.setenv("BOT_SPEAKER_CONFIRM_THRESHOLD", "1")
+    monkeypatch.setenv("VEXA_CSRC_INACTIVE_MS", "1200")
     monkeypatch.delenv("BOT_SPEAKER_SUBMIT_INTERVAL_SEC", raising=False)
     reg = default_registry()
     assert reg.get("meeting-bot").base_env == {
         "BOT_ALONE_SILENCE_WINDOW_MS": "60000",
         "BOT_SPEAKER_MIN_AUDIO_SEC": "1",
         "BOT_SPEAKER_CONFIRM_THRESHOLD": "1",
+        # the CSRC window override must actually reach the spawned bot's process env —
+        # capture-bridge guards the value; the kernel's job is only to carry it
+        "VEXA_CSRC_INACTIVE_MS": "1200",
     }
 
 

docs/changelog.d/1385-csrc-inactive-window.md

@@ -0,0 +1,7 @@
+- **Teams CSRC lanes stop fragmenting on natural speech pauses (#1383).** The transport sensor's
+  400 ms inactivity window spans a packet gap, not a speech pause — the median pause (p50 ≈ 550 ms)
+  already tripped a synthesized deactivation, splitting one speaker turn into ~1.7 lane activations.
+  The composition root now passes the measured 800 ms window (the knee: 1.13 activations/turn),
+  overridable via `VEXA_CSRC_INACTIVE_MS` where the bot process env is operator-controlled; an
+  unusable override (empty, NaN, below one 100 ms poll, above 10 s) warns and falls back to the
+  measured default instead of silently poisoning the sensor.