Step down when losing ground, not when slower than a constant

Build 117.

A campus M4 Pro sat in the red with nobody listening and no language being
translated — so transcription alone. Measured over SSH: sentences 4.2-4.7 s
against roughly 4-5 s of audio, queue 46 deep, lag 152 s and still climbing.

The mechanism built to catch exactly this could not fire. It triggers past 8 s,
on the reasoning that 8 s means slower than people talk. That held when a chunk
could be 8 s long; chunks are capped at HARD_CUT_MAX, 5.5 s. A 4.4 s sentence is
already losing and never comes near 8 s, so the ladder never stepped down and
the backlog grew without bound.

It now compares against the audio the chunk actually held. Two details are
measured rather than chosen: the margin is 0.85, because the Mini's 4.4 s
against 4.5 s was NOMINALLY faster than real time while still losing — the work
outside that timer (segmenting, the preview model sharing the GPU, the
scheduler) loses on its own. And it requires a backlog of five, so one slow
sentence on a healthy machine touches nothing. Checked against eight cases
including the Mini's real numbers and a healthy M2 Max.

The old absolute rule stays for a model that is genuinely hung.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

app/engine.js

@@ -1794,6 +1794,7 @@ async function pump() {
 
   state.busy = true;
   const started = Date.now();
+  let jobAudioMs = 0;
   try {
     /* job.at is when the audio ENDED. Live translation of stale audio is
        worthless, but the WORDS are not: however old the job, it still gets
@@ -1805,6 +1806,9 @@ async function pump() {
     /* Near-silence telemetry (log-only this release): a chunk whose average
        level barely clears the noise floor is hallucination fuel. */
     const audio = flatten(job.frames);
+    /* How much speech this chunk actually holds. The step-down below compares
+       against it, because "slow" only means anything relative to the audio. */
+    jobAudioMs = (audio.length / SAMPLE_RATE) * 1000;
     let sum = 0;
     for (let i = 0; i < audio.length; i += 4) sum += audio[i] * audio[i];
     const meanRms = Math.sqrt(sum / Math.ceil(audio.length / 4));
@@ -1908,11 +1912,43 @@ async function pump() {
      * talk, and the backlog only grows — step down before the service is lost
      * rather than after. Measured on the Intel laptop: 12-25 s, every sentence.
      */
-    if (took > 8000) {
+    /*
+     * SLOWER THAN THE AUDIO IT COVERS — not slower than some constant.
+     *
+     * This used to fire only past 8 s, on the reasoning that 8 s meant slower
+     * than people talk. That was true when a chunk could be 8 s long. Chunks
+     * are capped at HARD_CUT_MAX, 5.5 s, so a 4.5 s chunk taking 4.3 s is
+     * already losing ground and never came near the old threshold.
+     *
+     * MEASURED on a campus M4 Pro: every sentence 4.2-4.7 s against roughly
+     * 4-5 s of audio, a queue 46 deep and a lag of 152 s still climbing — with
+     * nobody listening and no language being translated, so this was
+     * transcription alone. The one mechanism built to catch exactly that could
+     * not fire, because 4.4 s is not 8 s.
+     *
+     * The queue test is what separates "this machine cannot keep up" from one
+     * slow sentence: a backlog that is three deep and being fed faster than it
+     * drains does not recover on its own.
+     */
+    /*
+     * 0.85 and a backlog of five, both measured rather than chosen.
+     *
+     * The Mini's sentences ran 4.4 s against 4.5 s of audio — nominally FASTER
+     * than real time, while the queue sat at 46 and the lag climbed past 150 s.
+     * Per-sentence time alone therefore misses it: the work outside this timer
+     * (segmenting, the preview model sharing the GPU, the scheduler) is enough
+     * to lose on its own. The margin catches "near enough to real time to be
+     * losing"; the queue depth is the proof that it actually is, and keeps one
+     * slow sentence on a healthy machine from touching anything.
+     */
+    const losingGround = jobAudioMs > 0 && took > jobAudioMs * 0.85 && queue.length >= 5;
+    if (took > 8000 || losingGround) {
       state.slowRun = (state.slowRun || 0) + 1;
       if (state.slowRun >= 3 && !respawning && lighterAsr(state.asrModelInUse || state.cfg.asrModel)) {
         state.slowRun = 0;
-        log('transcription is slower than people are speaking — stepping down to a lighter model', 'err');
+        log('transcription is slower than people are speaking (' + took + ' ms for ' +
+            Math.round(jobAudioMs) + ' ms of audio, ' + queue.length + ' waiting) — ' +
+            'stepping down to a lighter model', 'err');
         respawnWorker(state.asrWorker);
       }
     } else state.slowRun = 0;

app/package.json

@@ -2,7 +2,7 @@
   "name": "church-translate-engine",
   "productName": "Church Translate",
   "version": "3.23.2",
-  "buildNumber": 116,
+  "buildNumber": 117,
   "private": true,
   "description": "Live sermon translation — self-contained: captures audio, transcribes, translates, and serves the congregation page itself",
   "main": "main.js",
@@ -106,7 +106,7 @@
       ]
     },
     "asar": false,
-    "buildVersion": "116",
+    "buildVersion": "117",
     "win": {
       "target": [
         {