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

[jfw-ppi]

Delivers issue: #1383 — mechanism delivered; the issue's live leg (row A3) is now witnessed upstream by the author on a real Teams meeting (C0); the non-author sign-off is still open.

Contribution rights

• Independent: I created this contribution, or otherwise have the right to submit it
  under Apache-2.0, and it is not owned or controlled by an employer, client, or other entity.
• Employer/client authorization required: an employer, client, or other entity owns or
  may control this contribution. I am requesting Vexa's private corporate-authorization process.
• Unsure: I need a private rights review before merge.

Every commit carries my DCO Signed-off-by.

What & why

The transport sensor's CSRC_INACTIVE_MS = 400 spans a packet gap; in real Teams meetings it also does speech-pause duty, where it is too short: entry-timestamp staleness during natural pauses is p50 ≈ 550 ms, so the median pause trips a synthesized deactivation and one speech phase fragments. Replayed through the Teams CSRC channelizer, 400 ms gives 1.67 lane activations per real turn; 800 ms gives 1.13 (the knee) for ~1.8 pp more contamination; 1600 ms starts merging turns. Those numbers are a downstream measurement (#1383); this PR ports the mechanism.

• The composition root (capture-bridge.ts) passes 800 ms through the sensor's existing inactiveMs parameter. The sensor keeps its packet-gap default and now refuses a non-finite or sub-poll window from any caller — it is the point of introduction of the NaN failure mode (both liveness comparisons false, every source flipped on every poll).
• VEXA_CSRC_INACTIVE_MS overrides it per deployment: rendered onto the runtime by compose, helm (runtime.csrcInactiveMs) and lite, declared in the runtime's config.v1, forwarded to spawned bots by the kernel's tuning allowlist, documented in the configuration reference. Unset or blank ⇒ the default, silently — the templates render an unset knob as an empty string, and every other bot tuning knob reads it that way; a non-blank unusable value (NaN, below one 100 ms poll, above 10 s) warns and falls back.
• The two mixed-pipeline comments that stated the sensor's 400 ms as a fact now state the composed window.

Scope: fragmentation only. Lane contamination is out of scope (TEAMS_CSRC_CONTESTED_TRANSCRIPTS.md covers the adjacent contested-words problem).

Observation bundle

• C0 — the author's live witness (upstream, this head, real Teams) · a compose stack built from this tree (vexa-v012, PR bot image 5d493bfa972f, PR runtime image built from the same tree and pinned by its own tag; local CPU Whisper small via deploy/transcription; nothing external) in a real Teams meeting, one witness bot spawned through POST /bots, admitted by hand. Three rounds, each read off the witness's own <session>.csrc.jsonl sidecar and its [Csrc] csrc poll started (100ms, inactive after N ms) log line, with VEXA_CSRC_INACTIVE_MS set on the compose .env → runtime container → spawned bot env (verified by docker inspect at each hop):

  1. Synthetic speech, two speaker bots (Piper TTS), same ten-turn exchange at both windows: 1.00 activation per utterance at 800 ms and at 400 ms, one attributed transcript segment per utterance both times — no regression at 800 ms; synthetic speech never pauses long enough to discriminate the windows.

  2. One human reading ten scripted sentences with natural pauses, once per window: 400 ms → 13 activations, 7 re-activations within 0.6 s of the previous deactivation, 8 transcript segments with two garbled; 800 ms → 5 activations, 9 attributable segments. The reading cadence differed between the two readings, so this row shows the direction only.

  3. The same two recordings (the witness's own recordings of round 2, 74 s and 72 s), replayed through a speaker bot into both windows — cadence held constant:

     Recording (same audio in both columns)	400 ms: activations · re-activations < 0.8 s	800 ms: activations · re-activations < 0.8 s
     reading 1, 10 sentences	6 · 4	3 · 1
     reading 2, 10 sentences	9 · 8	4 · 3

  · concluded: on identical human audio the 800 ms window halves the transport activations for the same ten sentences and removes most sub-window re-activations, on both recordings; the composed window and the override reach the real bot on compose. Absolute counts are below the live reading because the replay re-encodes through a bot's virtual microphone; the ratio holds. Not claimed: the downstream per-turn numbers (C5), which need many speakers and hours; this is one voice, ten sentences, two recordings. Full timelines and tapes kept by the author; the recordings are the author's own voice and can be shared with a validator.

• C1 — sensor · pnpm --filter @vexa/mixed-capture-core test · green: 500 ms of silence held open inside a caller's 800 ms window, deactivation synthesized past it; inactiveMs of NaN, 50 and Infinity each fall back to 400 ms with one activation and no flip-flop, said through log.

• C2 — composition root · npx tsx src/csrc-wiring.test.ts on the built bundle, plain and with VEXA_CSRC_INACTIVE_MS=abc ambient · green both times: the real bridge hands the real sensor factory inactiveMs=800; garbage/zero/negative/sub-poll/over-ceiling fall back with one warning each; unset, blank and a good override are silent. Negative control: deleting the one inactiveMs: csrcInactiveMs, line from the page-side call turns 2 checks red — the two named RED-at-base.

• C3 — kernel + deploy surfaces · test_profiles.py + test_config_contract.py (22 passed); deploy/helm/tests/test_template.sh (gate:helm PASS); helm template --set runtime.csrcInactiveMs=1200 and docker compose config with the env set · the runtime carries VEXA_CSRC_INACTIVE_MS: "1200" on both.

• C4 — workspace · full @vexa/bot suite exit 0; gate:node (18 packages build + test) green; node scripts/gates.mjs all green except config-contract, which named the missing declaration — declared, that gate re-run green; the full sweep was not re-run after that one-line fix.

• C5 — measurements · downstream, prior to this port: staleness percentiles and the 400/800/1600 ms sweep over 5 × 30 min tapes. Not reproducible here — the tapes carry meeting audio, and the percentiles need raw poll samples a csrc.jsonl tape (post-window transitions) does not record.

Acceptance floor (the issue's numbering)

Row	Evidence
A1	Sensor default pinned (csrc-poll.test.ts); composition passes 800 (capture-bridge.ts CSRC_DEFAULT_INACTIVE_MS, asserted on the object that crosses to the sensor); override and unusable-window refusal pinned in csrc-poll.test.ts
A2	resolveCsrcInactiveMs + warning pinned in csrc-wiring.test.ts (floor/ceiling boundary-exact; blank silent); reachability pinned by test_profiles.py (kernel) and the helm/compose renders (C3)
A3	Witnessed upstream by the author (C0, round 3): same human audio, 400 → 800 ms halves activations on both recordings (6 → 3, 9 → 4). The downstream per-turn ratio (C5: 1.67 → 1.13) is not re-derived here. Non-author witness still open — see the validation request

Docs diff (D6c)

docs/changelog.d/1385-csrc-inactive-window.md (the mechanism proved; the reduction labelled a downstream measurement), the VEXA_CSRC_INACTIVE_MS row in docs/docs/configuration.mdx, runtime.csrcInactiveMs in values.yaml.

Security checks

No new dependencies, no lockfile change. The templates pass one env string through, the kernel copies it verbatim, the bot JSON.stringifys it in its one warning and hands the page a number. Nothing touches auth, secrets or the network.

Validation request

A non-author with Teams capture: reproduce C0 round 3 in about ten minutes — a stack from this head, one witness bot, and either read ten sentences with natural pauses yourself or replay any recording of natural speech through a second participant, once with VEXA_CSRC_INACTIVE_MS=400 on the runtime and once unset. Compare the activation count in the witness's csrc.jsonl (or the csrc-crossed counter in its log) for the same audio; the author's two recordings are available on request. Deployments validated in this PR: compose, live (C0); helm and lite: template render only (C3), no live run — honestly unclaimed.

Era

Based on and verified against main@3f5c3c03 (the v0.12.26 train).

Authorship

Sole author: the human submitting this. No agent co-author trailers (D13). Tooling disclosure: port and verification assisted by Claude Code; the measurements are from live captures in a downstream deployment.

[github-actions]

👋 Thanks for opening your first PR to Vexa, @jfw-ppi!

Highly recommended (not required): hop into our Discord
and tell us in a sentence or two what this change does and why. It helps us review
your value bundle faster and connects you with the reporter and maintainers.

Your PR is judged on its evidence — the observation bundle + the diff — not on whether you
show up. But showing up makes the whole thing smoother. See
the delivery guide for how a PR reaches merge.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 2c18d14099

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[github-actions]

🃏 Merge card — #1385

check		what it needs
Value	❌	missing state: value-signed (the value sign-off)
Diff	❌	no non-author approval on the current head sha (a new push dismisses a stale approval)

Not mergeable yet — every row above must be accepted before merge (choke point 1). Fill in what's ❌ above, then this clears automatically.

_{How a PR reaches merge: the merge bar.}

[jfw-ppi]

Self-review against main@3f5c3c03 (head b61e54ee at the time). Findings addressed in 926b8d0f; verification is in the PR description.