…boundary (#24)

The streaming C API conflated end-of-utterance and backchannel: the feed
eou_out flag and the JSON "eou" field were 1 for either token, so a voice
agent could not tell "the user yielded the turn" (<EOU>, respond) from
"the user acknowledged while listening" (<EOB>, do not treat as a turn).

- parakeet_capi_stream_feed's *eou_out is now an event bitmask
  (PARAKEET_EVENT_EOU | PARAKEET_EVENT_EOB) instead of an any-event 0/1
- the streaming JSON documents carry separate "eou"/"eob" 0/1 flags plus
  an "events" array ({"type":"eou"|"eob","frame","t"}) with per-event
  timestamps
- new parakeet_capi_stream_drain_events / parakeet_capi_free_events
  expose typed per-event records (token, is_eob, encoder_frame,
  time_sec) on the flat API; the queue is shared with the JSON entry
  points
- StreamingSession gains a non-draining events() peek; the C layer
  watermarks the queue to attribute events to one feed pass (a feed can
  span several encoder chunks, which last_chunk_had_eou() cannot express)
- decoder reset-on-<EOU>/<EOB> is deliberately unchanged (matches NeMo's
  reference driver); the distinction is a downstream turn-taking signal
- tests: both C-API stream tests gain a two-utterance-clip phase (shared
  builder in tests/stream_clips.hpp, same construction as
  gen_stream_reset_baseline.py) asserting the typed mid-stream <EOU> and
  exact flag/mask agreement with the drained queue

Verified against mudler/parakeet-cpp-gguf realtime_eou_120m-v1-f16: all
61 ctest targets pass; the two-utterance clip yields the mid-stream
<EOU> at t=7.84s with mask=PARAKEET_EVENT_EOU.

Assisted-by: Claude:claude-fable-5 [Claude Code]

ggml leaves GGML_CUDA_GRAPHS off by default. Turning it on lets the CUDA
backend capture and replay the compute graph, a small but free speedup
(about 1% measured on a GB10: 1.4% on tdt-1.1b, 0.4% on tdt-0.6b-v3, and
never negative across interleaved runs).

Enable it whenever we forward CUDA so every CUDA build (docker, release,
local) inherits it. The runtime kill-switch GGML_CUDA_DISABLE_GRAPHS=1
still disables it for A/B testing.

Co-authored-by: Ettore Di Giacinto <[email protected]>
Co-authored-by: Claude Opus 4.8 <[email protected]>

* feat(server): OpenAI-compatible transcription server example

Squashed rebase of the worktree-openai-server branch (PR #8) onto current
master. Adds examples/server (httplib-based POST /v1/audio/transcriptions),
model resolver/fetcher, OpenAI response formatter, and unit + e2e tests.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

* feat(docker): publish a dedicated parakeet-server image

Split the Dockerfile into a shared build stage plus two runtime targets:
runtime (cli, default, unchanged) and runtime-server (entrypoint
parakeet-server --host 0.0.0.0, EXPOSE 8080, curl for alias fetch). The
docker workflow now builds and publishes both ghcr images, cli and server,
for each (variant, arch); the server build reuses the cli build stage so
ggml compiles once per job.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

* docs: document the OpenAI server and docker images; point to LocalAI for production

Add an OpenAI-compatible server section to the main README (build, curl and
OpenAI-client usage) and extend the Docker section to cover the new
parakeet.cpp-server image alongside the cli. Note LocalAI as the production
path in both the main README and examples/server/README.md.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

* ci: add server-e2e job driving the real parakeet-server over HTTP

Builds parakeet-server and runs tests/server_e2e.sh (PARAKEET_SERVER_E2E=1):
fetches the ~125 MB tdt_ctc-110m-q4_k model by alias, starts the server, and
hits POST /v1/audio/transcriptions with a real WAV in json/text/verbose_json
(plus word timestamps), checking the transcription and the 400 paths. Runs on
pull_request and workflow_dispatch, like closed-loop; no NeMo/Python venv.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

---------

Co-authored-by: Claude Opus 4.8 (1M context) <[email protected]>