Reuse the language-detection encoder pass in long-form decodes

Whisper's generate() keeps only forward_params from its kwargs, and that
list omits encoder_outputs, so the precomputed encoder output was dropped
and auto-detect cost three encoder passes for a one-chunk clip. The
real-model test now pins one encoder pass per decoded chunk.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

docs/plans/001-3-longform-pipeline.md

@@ -207,7 +207,7 @@ normalize), `scripts/measure-longform.ts`.
 - [ ] Run under the lock, old/new interleaved: baseline; long-form guard off; guard variants
   (`no_repeat_ngram_size 3/5`, `repetition_penalty 1.1`); fallback on if loops remain; VAD
   `min_silence 2000` vs `160`. Choose defaults by the rollout rule; record the table in D-WS3a.
-- [ ] Commit "Add the synthetic meeting fixture and long-form measurement script".
+- [x] Commit "Add the synthetic meeting fixture and long-form measurement script".
 
 ### Task 7: Worker, protocol, engine and UI wiring
 

src/worker/longform/transformers.ts

@@ -211,6 +211,14 @@ export function createWhisperDeps(
     };
   };
 
+  // generate() keeps only the model's `forward_params` from its kwargs, and
+  // Whisper's list omits `encoder_outputs` — so a precomputed encoder output
+  // was silently dropped and the encoder ran again. Admitting it on this
+  // instance lets generate() skip the encoder (its own `"encoder_outputs" in
+  // model_inputs` check) when detection already encoded the chunk.
+  const params = (asr.model as unknown as { forward_params: string[] }).forward_params;
+  if (!params.includes("encoder_outputs")) params.push("encoder_outputs");
+
   /** Encoder output of the chunk language detection just ran on. */
   let encoded: { audio: Float32Array; hidden: Tensor } | null = null;
   const decodeReusingEncoder = (audio: Float32Array, request: DecodeRequest) => {

tests/live-real-model.test.ts

@@ -352,6 +352,17 @@ describe("long-form file pipeline accuracy (real model, real recordings)", () =>
             // "Auto-detect" in the UI: the chunked path forces English here
             // (Transformers.js 4.2 has no detection); long-form detects once
             // on the first speech chunk and pins it (D-WS3a).
+            // Detection must not cost an extra encoder pass: the first
+            // chunk's decode reuses the encoder output detection computed.
+            const model = asr.model as unknown as Record<string, unknown>;
+            const encode = model._prepare_encoder_decoder_kwargs_for_generation as (
+              ...args: unknown[]
+            ) => Promise<unknown>;
+            let encoderPasses = 0;
+            model._prepare_encoder_decoder_kwargs_for_generation = (...args: unknown[]) => {
+              encoderPasses += 1;
+              return encode.apply(model, args);
+            };
             const auto = await transcribeLongForm(
               audio,
               {
@@ -359,8 +370,12 @@ describe("long-form file pipeline accuracy (real model, real recordings)", () =>
                 ...createWhisperDeps(asr, { multilingual: true }),
               },
               { ...LONG_FORM_DEFAULTS, task: "transcribe" },
-            );
+            ).finally(() => {
+              delete model._prepare_encoder_decoder_kwargs_for_generation; // back to the prototype's
+            });
             expect(auto.language).toBe(clip.language);
+            // One encoder pass per decoded chunk (fallback is off by default).
+            expect(encoderPasses).toBe(auto.decodeCalls);
             expect(
               wordErrorRate(normalize(reference), normalize(auto.text)),
             ).toBeLessThanOrEqual(MAX_WER_BY_CLIP[clip.base] ?? MAX_WER);