Measure Jim's audio and keep his volume

docs/latency-observability.md

@@ -169,6 +169,40 @@ out, and input transport before the agent on the way in, are outside this
 record. A browser-perceived figure needs a matched audio measurement in the
 browser.
 
+## Browser audio smoke evidence
+
+The credentialed `voice` and `barge` browser-check flows also capture
+`jimAudio`: received byte/packet deltas, peak decoded RMS, playback-blocked
+state and a bounded result per observed audio subscription. The harness uses
+LiveKit's public track-specific `getRTCStatsReport()` and an analyser on the
+same native track attached as Jim. It records no device labels or audio
+samples.
+
+Sampling waits up to 120 seconds for Speaking, then samples at 100 ms
+intervals for at most 10 seconds, stopping early once both counters grow and
+RMS exceeds 0.001. Counter resets and receiver changes start a new baseline;
+different track generations never share deltas. At most 64 subscriptions are
+sampled; exceeding that limit is an unavailable observation, not a healthy
+verdict.
+
+`unavailable` means there were insufficient counter pairs or decoded samples,
+no attached Jim track, no Speaking interval, cancellation, a sampling error or
+a subscription limit. `transport-silent` means bytes or packets, either one,
+did not grow across the observed counter pairs, and `decoded-silent` means both
+grew but decoded RMS stayed at or below the floor. These are separate failures.
+The capture retains failed observations, and each resubscription that keeps its
+track through its sampling window must pass independently. A generation whose
+track is replaced, unsubscribed or disconnected before Speaking or within its
+first second of it, before both counters grew and RMS exceeded the floor, is
+marked `superseded`, since losing the track cut its window short and a fresh
+analyser first reads near zero; it cannot fail an otherwise healthy
+replacement, and the last generation must pass. A generation that sampled
+longer is judged on its evidence even if its track then ends. Unmeasured
+counters and RMS are `null`, not observed zeroes. This measures transport and
+decoded signal during Speaking, not output-device audibility, physical level or
+per-word delivery. The RMS floor is a smoke-check threshold, not a calibrated
+speech detector. Local loopback tests do not replace the isolated provider run.
+
 ## Validation
 
 The credential-free gate covers the trace lifecycle, the waterfall and the

docs/meet-tab-audio-manual-check.md

@@ -37,20 +37,25 @@ re-vendored, or before relying on the mode for a real interview.
    confirm Jim moves to it. This step is what catches a broken device list or a
    preference that fails to persist, so do not skip it when Jim is already
    audible.
-3. In Meet, click Present now.
-4. Choose "A tab".
-5. Select the CodeTrial tab.
-6. Enable "Share tab audio" before clicking Share.
-7. Confirm Meet's microphone is still the candidate microphone, not "None" and
+3. Open "Audio controls" on the stage. Jim's volume and mute set the level of
+   the tab audio Meet captures, and the volume is saved across sessions, so a
+   low level left from an earlier run reaches the interviewer as a quiet Jim.
+   Press Reset volume, and confirm the slider reads 100% and "Mute Jim" is not
+   pressed.
+4. In Meet, click Present now.
+5. Choose "A tab".
+6. Select the CodeTrial tab.
+7. Enable "Share tab audio" before clicking Share.
+8. Confirm Meet's microphone is still the candidate microphone, not "None" and
    not a virtual device.
-8. Record 10 seconds in Meet while Jim is speaking and the candidate says one
+9. Record 10 seconds in Meet while Jim is speaking and the candidate says one
    sentence.
 
 ## Pass criteria
 
 All five must hold:
 
-- The recording contains Jim's voice.
+- The recording contains Jim's voice at the level set in step 3.
 - The recording contains the CodeTrial tab video.
 - The recording contains the candidate's voice.
 - Meet's selected microphone is still the candidate's physical microphone at

scripts/browser-check.cjs

@@ -10,6 +10,7 @@
 const { chromium } = require(process.env.PLAYWRIGHT_PATH);
 const assert = require("node:assert/strict");
 const fs = require("fs");
+const { mediaProbeSource, assertJimAudio } = require("./media-probe.cjs");
 const { installBargeReplay } = require("./barge-replay.cjs");
 const {
   FloorCapture,
@@ -793,6 +794,8 @@ async function isolateRustAgent(
   try {
     const page = await browser.newPage();
     let bargeReplay = null;
+    let jimAudio = null;
+    let probingAudio = false;
     if (process.env.BROWSER_CHECK_FLOW === "barge") {
       await page.addInitScript(
         installBargeReplay,
@@ -2310,6 +2313,10 @@ PatternLexicon.prototype.search = function(word) {
     }
 
     await page.goto(interviewUrl(PROBLEM_IDS.twoSum));
+    probingAudio = credentialed && ["voice", "barge"].includes(flow);
+    if (probingAudio) {
+      await page.evaluate(mediaProbeSource());
+    }
     await clearMediaGate(page);
     if (credentialed && !rustAgentIdentity) {
       roomName = await page
@@ -2478,6 +2485,14 @@ PatternLexicon.prototype.search = function(word) {
             );
           }
         }
+        if (probingAudio) {
+          jimAudio = await page.evaluate(() =>
+            globalThis.__finishJimAudioProbe(),
+          );
+          if (process.env.BROWSER_CHECK_CAPTURE)
+            writeIncompleteCapture({ jimAudio, bargeReplay });
+          assertJimAudio(jimAudio);
+        }
         if (credentialed && flow === "barge") {
           if (!floorLogPath)
             throw new Error(
@@ -2543,6 +2558,7 @@ PatternLexicon.prototype.search = function(word) {
                   ? readFloorRecords(floorLogPath)
                   : [],
                 bargeReplay,
+                jimAudio,
                 latencyLog: (agentOutput.join("") + "\n" + readServerLog())
                   .split("\n")
                   .filter((line) => line.includes("codetrial latency "))
@@ -2557,10 +2573,16 @@ PatternLexicon.prototype.search = function(word) {
       })(),
       agentFailure ?? never,
     ]).catch(async (error) => {
-      if (process.env.BROWSER_CHECK_CAPTURE && floorLogPath)
+      if (probingAudio && !jimAudio) {
+        jimAudio = await page
+          .evaluate(() => globalThis.__finishJimAudioProbe(true))
+          .catch(() => null);
+      }
+      if (process.env.BROWSER_CHECK_CAPTURE && (floorLogPath || probingAudio))
         writeIncompleteCapture({
-          floorRecords: readFloorRecords(floorLogPath),
+          floorRecords: floorLogPath ? readFloorRecords(floorLogPath) : [],
           bargeReplay,
+          jimAudio,
         });
       console.error(
         redact(