server: resolve the model contract once per model, not per request

[XythQ]

Problem

ServerState::build_speech_request calls model_accepts_request_option() on every
POST /v1/audio/speech. That resolves the model contract through
engine::model_spec::model_contract(family), which reads the model's embedded spec with
gguf_init_from_file() on the full GGUF.

The resolution cache is thread_local (src/framework/model_spec/package.cpp:22-26), and
serve_http handles each request on a fresh detached thread, so the cache is cold every time. Each
request re-opens and re-parses the whole model file.

With a 298 MB GGUF this adds a fixed ~900 ms to every request, independent of backend, text
length and audio duration. It is invisible in X-AudioCPP-Wall-Ms because run_model starts its
clock after build_speech_request, and it does not affect audiocpp_cli, which resolves the spec
once on a single thread.

Fix

Resolve the option once per model instead of once per request:

• LoadedModel gains bool accepts_reference_text (defaults to true, the existing permissive
  behaviour for models with no contract).
• refresh_model_option_flags() sets it in make_model, and again in handle_model_load when a
  model is reconfigured, so dynamic model management stays correct.
• build_speech_request reads the cached flag.

A contract that fails to resolve is caught at registration, logged to stderr, and falls back to
the permissive default — same behaviour as before, but reported once instead of throwing per
request.

The one-off resolution cost moves to registration, where the model load already performs a
gguf_init_from_file.

Measurements

v0.7.0, Windows x64, Vulkan backend, supertonic-3-f16.gguf (298 MB), same server.json, same
request:

	ttfb	X-AudioCPP-Wall-Ms
before	0.956–1.153 s	44–176 ms
after	0.045–0.049 s	44–50 ms

Also reproduced before the fix on --backend cpu (ttfb 1.268–1.414 s against 341–362 ms reported),
confirming it is not backend-specific.

Response bytes are unchanged: the same request returns Content-Length 114090 with an identical
MD5 (1514cf1fcf6bfc4e865a1cc24d5b8948) before and after.

Notes

• Long utterances now scale with synthesis time rather than sitting on a fixed floor.
• This caches the reference_text lookup, the only contract option the speech path uses today.
  Any other per-request contract lookup added later would hit the same cold thread_local path, so
  it may be worth resolving the contract once per model more generally.

[0xShug0]

@XythQ Thanks, this is a useful improvement! One small issue before merge: the new fallback swallows real model-spec errors.

With an invalid spec override, startup logs the real error but continues permissively:

[server] model 'bad-contract-smoke': reference_text option check failed (model spec '/tmp/audiocpp_pr328_bad_spec.json'.schema_version: expected 1); assuming the option is accepted
audiocpp_server listening on http://127.0.0.1:18090

Then the request fails later with an unrelated error:

HTTP/1.1 500 Internal Server Error
{"error":{"message":"model path does not exist: /tmp/nonexistent-supertonic-smoke.gguf","type":"server_error"}}

A better approach may be to remove the broad catch and let invalid specs fail during registration or load.

[XythQ]

Removed rather than narrowed — model_accepts_request_option already returns true for a model with no contract and rethrows the rest (runtime.cpp:107), so the catch only ever covered real misconfigurations. An invalid spec now throws at registration and reaches main.cpp:235 as audiocpp_server failed: <error>, exit 1; a throw during dynamic load hits the existing handler at http.cpp:777.

One consequence: a configured-but-absent override file now fails loudly. Intentional, but I didn't build to reproduce your case — the log line traces statically to fail() at schema.cpp:16-18.

[0xShug0]

@XythQ Thanks! PR merged 🚀