[boydjohnson]

Milestone 2. transcribe reaches rkwhisperd through rkwhisper-client in the daemon, and POST /v1/audio/transcriptions serves it in the frontend.

The path
The frontend takes the multipart upload, decodes the container with symphonia, resamples to 16 kHz mono s16le with rubato, and hands the daemon a stream rather than a buffer, so rkwhisperd starts on the first second of a recording while the rest is still being decoded and a long clip is never held in memory as PCM. The daemon opens an rkwhisper-client session and maps rkwhisperd's responses onto this protocol's events. All five response formats are served: json, text, srt, vtt and verbose_json.

Transcription goes through the daemon rather than straight from the frontend, because rkwhisper-client needs rkwhisperd's Unix socket and shared-memory audio ring — a frontend using it directly would have to run on the board. This way the frontend speaks one protocol and /v1/models has one source.

Also here
Model state for rkwhisper models. They load nothing locally, so they had no loader thread to report progress and sat unavailable permanently, which refused every transcribe call at validation. The probe is a session opened and dropped: more than reaching the socket, since the handshake names the model; less than a transcription, since no audio follows.
A body limit that was hiding a 25 MB cap. Axum's DefaultBodyLimit is 2 MB, so the endpoint's documented OpenAI limit could never fire and a 3 MB recording came back as "Error parsing multipart/form-data request". An oversized upload is now a 413 that says so.
A resampler count fix. rubato reports a whole block consumed even for a partial one, so clips looked longer than they were and the padding the count exists to drop went through instead. Lengths are exact now, and the tests assert them exactly rather than within a tolerance wide enough to hide it.
SDK coverage. tests/sdk_conformance.py drives the official SDK against the endpoint: the four response formats, a 44.1 kHz upload that exercises the resampler, and an upload that is not audio. It builds its own WAV, so it still needs no fixture beside it.
Verification
cargo fmt, cargo clippy --workspace --all-targets -- -D warnings and cargo test --workspace --all-targets (207 tests) are clean, and the release build cross-compiles for aarch64-unknown-linux-gnu. The frontend's tests run against a fake daemon, so they need no hardware.

Exercised end to end on an RK3588 against rkwhisperd with whisper-small-30s: a ten-minute clip returns a transcript, timed segments and verbose_json.

Known, and not addressed here
Transcript coverage on long audio depends on rkwhisper's own decode settings, which this PR does not expose. rkwhisperd's ClientHello defaults are what it ships with — beam 5, timestamps on, 128 new tokens per window. Measured on a 9.6-minute clip, dropping to greedy decoding (beam_size = 1) covered only 14% of the audio: a degenerate pass trips rkwhisper's repetition guard, and the stale timestamp it last emitted drives the seek loop past whole windows. So beam_size, notimestamps and max_new_tokens stay at rkwhisper's defaults and are not configurable yet. Whether they become per-model config or per-request fields is worth settling separately, with numbers.

verbose_json's avg_logprob, compression_ratio, no_speech_prob and tokens are zeros and empties standing in for values rkwhisper does not report. They are present because the SDK's schema requires the fields.

🤖 Generated with Claude Code