QVAC-24488 feat[api]: close the SDK gaps against @qvac/bci-whispercpp 0.9.1

[RamazTs]

Stacked on #4564. Based on feat/QVAC-24486-sdk-asr-ggml-gaps so the diff shows only the BCI work — #4564 renames WhisperAddonSegment → AsrAddonSegment and touches ops/bci-transcribe.ts, which would otherwise conflict. GitHub retargets this to main once #4564 merges.

🎯 What problem does this PR solve?

Gap audit of @qvac/bci-whispercpp 0.9.1 against the SDK surface. Already pinned ^0.9.1, so parity work only — no version bump.

• Streaming segments lost windowStartTimestep. lib/stream.js:148 attaches it to every segment emitted in emit: 'delta' mode, and the addon documents exactly why: a segment's own timestamps are window-local, so this is the only thing that maps them onto the stream timeline. It appeared nowhere in the SDK, leaving delta-mode output unplaceable.
• Six emitted stats fields never reached callers: processCalls, totalTime, totalWallMs, whisperSampleMs, whisperBatchdMs, whisperPromptMs.
• Five mapped fields the addon never emits: audioDurationMs, realTimeFactor, encoderMs, decoderMs, melSpecMs were declared on BciAddonResponse and conditionally spread, but BCIModel.cpp emits none of them — they belong to the asr-ggml engines. Harmless at runtime, but the interface promised a shape that cannot arrive.
• No backend diagnostics, where the asr-ggml ops now have them.
• detect_language was documented as rejected by the addon. It is in the addon's own validWhisperParams; the wording was inherited from the asr-ggml schema.

📝 How does it solve it?

• windowStartTimestep is carried on the shared segment shape, and only when the engine sent it — batch segments do not gain a key that would be meaningless.
• The stats interface is rewritten to the set BCIModel.cpp actually emits, and all six missing fields are mapped. The five phantom fields are gone.
• Backend diagnostics ride the batch terminal frame, reusing the buildAsrBackendDiagnostics mapping from QVAC-24486 feat[api]: close the SDK gaps against @qvac/asr-ggml 0.5.3 #4564, and the frame attaches the profiling symbol so they land on the profiling event's backend. The streaming path is deliberately untouched: the addon does not populate response.stats for streams, so there is no verdict to derive.
• detect_language's description now matches the validator.

Where the stats and diagnostics reach. They are on the batch bciTranscribe terminal frame, so RPC consumers see them (Python's bci_transcribe yields that frame) and profiling picks up the diagnostics. The JS convenience functions do not surface them: bciTranscribe() resolves to text or segments only, and neither BCI stream session (JS or Python) has a stats field. Exposing them there changes the public return shapes, so it is left as a follow-up rather than done here.

Verified already at parity and left unchanged: whisperConfig 16/16, contextParams 4/4, miscConfig 1/1, bciConfig 1/1 (day_idx), and the per-call stream options windowTimesteps / hopTimesteps / emit (including the hop < window refine that mirrors the addon's own check).

🧪 How was it tested?

• inference: 146/146 suites green.
• SDK typecheck, SDK test-types, and lunte on both packages: clean.
• contract:export --check and sdk-python generate.py --check: both exit 0, with both zod copies pinned to the version CI resolves.
• New coverage in bci-schemas.test.ts: windowStartTimestep survives the mapper and is absent on batch segments; the terminal frame accepts a diagnostics payload; the full emitted stats surface round-trips.
• New bci-reachability.test.ts drives the real handlers rather than the schemas: the batch terminal frame feeds buildOperationEvent's backend, and delta-mode segments keep windowStartTimestep through the handler output. Each test fails against the code it guards (the handler without the symbol attach, and the mapper without the field).

🔌 API Changes

// Delta-mode segments can now be placed on the stream timeline.
for await (const seg of session) {
  if (seg.windowStartTimestep !== undefined) {
    const absoluteStartMs = seg.startMs + timestepsToMs(seg.windowStartTimestep)
  }
}

# The stats the native model actually reports, on the batch terminal frame.
async for frame in bci_transcribe(transport, request):
    if frame.done:
        print(frame.stats.total_wall_ms, frame.stats.process_calls, frame.diagnostics)

📋 Follow-ups

• BCI stats in the JS API. bciTranscribe() drops the batch terminal frame's stats and diagnostics. Surfacing them means widening its return type, which is an API decision of its own. The stream sessions have nothing to surface until the addon reports stats for streams.
• computeWER is exported by the addon and surfaced nowhere. Left out on purpose: it is a scoring helper rather than part of the inference path, so exposing it is an API-design call rather than a parity fix.
• reload() is not a BCI gap. It exists on BCIInterface — the native binding wrapper exposed as the addon escape hatch — but the public BCIWhispercpp class has none, and the SDK's reload path calls model.reload(...). Widening the whisper-only reload union would not reach BCI.

[github-actions]

License compliance — clean

No new dependency license findings in this PR.

Warn-only (shadow) mode — this check does not block merges yet.

Updated automatically by the canonical license compliance workflow.

NOTICE presence (advisory)

Missing NOTICE (advisory, does not block):

• ./docs/website
• ./packages/fabric/test/integration
• ./packages/llm-llamacpp/benchmarks/server
• ./packages/llm-llamacpp/benchmarks/performance
• ./packages/inference-addon-cpp/mobile
• ./packages/asr-ggml/benchmarks/server
• ./packages/embed-llamacpp/benchmarks/server
• ./packages/embed-llamacpp/benchmarks/performance
• ./packages/sdk/e2e
• ./packages/vla-ggml/sim/server
• ./.github/actions/release-merge-guard

The base branch was changed.