[LingXuanYin]

Summary

Adds scene3 ("Scene 3"), a text-to-speech model with inline delivery cues and one-reference cloning. Rebased onto main after #56 / #83 / #84; qwen-audio-3.1-tts-flash keeps its validation and request body.

 built-in models / models/*.yaml
+  scene3 (category audio, adapter openai.audioSpeech)
 openai.audioSpeech adapter
   qwen-audio-3.1-tts-flash: unchanged
+  scene3: one text block; at most one HTTP(S) reference; audio meta.text = that reference's transcript
 README
+  scene3 in the model list, a note in the model-choice bullets (used when named), one example

Request body sent to /v1/audio/speech:

no reference   { model: "scene3", input }
one reference  { model: "scene3", input, metadata: { references: [{ url, text? }] } }

input is the text block as given; [bracketed] cues such as [sighs] stay in it. The URL and transcript are trimmed, and text is sent only when a transcript was given. Request-level meta (for example voice_prompt) and any audio meta other than text are rejected before the request is sent.

Evidence

• Before: on 1e3b000 with only the new tests, 15 fail: the model does not exist. With the declaration but no adapter branch, 9 fail with Unsupported audio speech model: scene3: the three request tests, the four adapter-only rules (request meta, audio meta key, blank or non-string meta.text), the overridden-declaration limit, and the built-in example check.
  After: rebased onto 4b728ac as aabe5a4, exported without line-ending conversion, in node:24-slim: pnpm lint checks 72 files with no errors, typecheck passes, vitest 411 passed (19 files; main alone has 397), build completes.
• Rebase conflicts were only the neighbouring qwen lines (README model list and choice notes, the adapter's model constants, the built-in speech model array); the scene3 lines are unchanged.
• Live tests: none added and none run. The gateway does not serve scene3 yet; live cases come with the end-to-end check once it does.
• Added lines, file names and commit messages were scanned for internal names: clean.

Merge Danger

Door: two-way for the code; a published version stays.

Blast Radius: none for existing callers. The change only adds a model id; the qwen path is untouched.

The version bump comes in a separate PR (0.2.2), as usual. Dependents still on ^0.1.x get the same change from release/0.1 (#86).