Read the timing keys the harness actually emits

The page showed "926.2 ms · retrieval + guardrails + reader · against a
200 ms budget" in red. 877.6 ms of that was speech synthesis.

Both pages read t.tts_ms and t.asr_ms. The harness names every timing
after its stage, so the keys are speak_ms and transcribe_ms and the two
the pages wanted have never existed. A missing key reads as absent rather
than as an error, so `spoken` was always false: the branch that appends
"+ speech synthesis" to the label, and the line that breaks synthesis out
of the total, were both unreachable. The label was honest about a value it
was never given.

Fixed in web/index.html and in serve.py's PAGE. The split line now also
leads with the graded path on its own -- 48.5 ms here against 926.2 end to
end -- because one number dominated by synthesis tells a reader nothing
about retrieval.

test_pages_read_timing_keys_the_harness_emits pins the contract: every
t.<x>_ms a page reads must be a key the harness emits, with the stage
names parsed from harness.py rather than duplicated. Verified non-vacuous
and verified to reject tts_ms and asr_ms.

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

src/serve.py

@@ -496,10 +496,10 @@ function render(d){
       '</div>' : '')+
     '<div class="lbl">answer</div><div class="ans">'+(d.answer||"—")+'</div>'+
     '<div class="t">'+
-      (t.asr_ms?'<span>ASR <b>'+t.asr_ms+' ms</b></span>':'')+
+      (t.transcribe_ms?'<span>ASR <b>'+t.transcribe_ms+' ms</b></span>':'')+
       '<span>retrieve <b>'+ms(t.retrieve_ms)+'</b></span>'+
       '<span>read <b>'+ms(t.read_ms)+'</b></span>'+
-      (t.tts_ms?'<span>TTS <b>'+t.tts_ms+' ms</b></span>':'')+
+      (t.speak_ms?'<span>TTS <b>'+t.speak_ms+' ms</b></span>':'')+
       '<span>total <b>'+ms(t.total_ms)+'</b></span>'+
       '<span>conf <b>'+d.confidence+'</b></span>'+
     '</div>'+

tests/test_deploy.py

@@ -365,6 +365,28 @@ def test_gradio_ssr_is_off_before_the_import():
     assert "ssr_mode=False" in src, "launch/mount must also pass ssr_mode"
 
 
+def test_pages_read_timing_keys_the_harness_emits():
+    """The harness names every timing after its stage: speak_ms, transcribe_ms.
+    Both pages read tts_ms and asr_ms, which never exist -- so the budget line
+    said "retrieval + guardrails + reader" over a number carrying 877 ms of
+    speech synthesis. A missing key reads as absent, not as an error, which is
+    why it survived. Any t.<x>_ms a page reads must be a key harness emits."""
+    import re
+
+    # Stage names are declared inline in _stages; read them from the source so
+    # this test needs no built index to run.
+    hsrc = (REPO / "src" / "harness.py").read_text(encoding="utf-8")
+    stages = re.findall(r'Stage\("([a-z_]+)"', hsrc)
+    assert len(stages) >= 7, f"expected 7 stages, found {stages}"
+    allowed = {n + "_ms" for n in stages} | {"total_ms", "pipeline_ms", "budget_ms"}
+
+    for rel in ("web/index.html", "src/serve.py"):
+        text = (REPO / rel).read_text(encoding="utf-8")
+        used = set(re.findall(r"t\.([a-z_]+_ms)", text))
+        unknown = used - allowed
+        assert not unknown, f"{rel} reads timing keys the harness never emits: {sorted(unknown)}"
+
+
 if __name__ == "__main__":
     tests = [v for k, v in sorted(globals().items()) if k.startswith("test_")]
     for t in tests:

web/index.html

@@ -884,14 +884,19 @@ function renderBudget(t){
      transcribe_ms. Speech synthesis is therefore INSIDE the budget, and a voice
      request legitimately blows it where a text request does not. Saying
      otherwise on the page would be the pleasant version, not the true one. */
-  const spoken = (t.tts_ms ?? 0) > 0;
+  const spoken = (t.speak_ms ?? 0) > 0;
   $("#budgetLbl").innerHTML = spoken
     ? "retrieval + guardrails + reader <b>+ speech synthesis</b>,<br>against a 200 ms budget"
     : "retrieval + guardrails + reader,<br>against a 200 ms budget";
 
   const parts = [];
-  if(t.asr_ms) parts.push(`speech-to-text, excluded <b>${t.asr_ms.toFixed(0)} ms</b>`);
-  if(spoken)   parts.push(`of which synthesis <b>${t.tts_ms.toFixed(0)} ms</b>`);
+  // The graded path on its own. Without it the page shows one number, dominated
+  // by synthesis, and a reader cannot tell retrieval from text-to-speech.
+  if(spoken) parts.push(
+    `retrieval + guardrails + reader <b>${(ms - t.speak_ms).toFixed(1)} ms</b>`);
+  if(t.transcribe_ms) parts.push(
+    `speech-to-text, excluded <b>${t.transcribe_ms.toFixed(0)} ms</b>`);
+  if(spoken) parts.push(`of which synthesis <b>${t.speak_ms.toFixed(0)} ms</b>`);
   parts.push(`end to end <b>${(t.total_ms ?? ms).toFixed(0)} ms</b>`);
   $("#voiceSplit").innerHTML = parts.map(p => `<span>${p}</span>`).join("");
 }