feat: carry an audio prompt's transcript text

[wavekat-eason]

Why

Today an audio prompt in a flow document is a bare reference — prompt: { audio: vprompt_… }. The words the clip actually speaks are dropped: they exist where the clip was authored, but nowhere in the document. So a read-only viewer can play a prompt clip yet cannot show what it says — there's no transcript to display alongside playback.

The text already exists upstream (it's what the clip was synthesized from). This change gives the document a place to carry it.

What

Add an optional transcript to the Prompt audio variant — the words the clip was synthesized from, carried for display (what a caller hears) and for traces.

• Schema (schema/flow.v1.schema.json): Prompt audio object gains an optional transcript. Regenerated the TS model + bundled schema via pnpm gen; the crate-local copy syncs through build.rs, and typify emits Prompt::Audio { audio, transcript: Option<String> } with #[serde(default, skip_serializing_if)] so it round-trips and stays absent when unset.
• Helpers (model_ext.rs, model.ts): new Prompt::transcript() / promptTranscript() return the spoken words for either prompt kind (a text prompt is its own transcript; an audio prompt returns its transcript, or None when omitted).
  • as_text() / promptText() are unchanged — they still mean "text to synthesize" and return None for an audio prompt. Playback branching and the 2000-char length cap are deliberately untouched; an audio transcript is display text, not something to synthesize or bound.

Why transcript, not text

promptText() deliberately returns null for an audio prompt (it means "text to synthesize"). A field also named text would sit right beside a helper that returns null for it — promptText(p) === null while p.text is a non-empty string. transcript removes that contradiction, matches the accessor name, and reads as the user-facing concept ("the transcript of what the clip says"). It intentionally does not mirror the platform's source_text column name, which describes provenance rather than the viewer-facing meaning.

• Conformance: a new valid case (audio-prompt-transcript) exercises transcript present on one node and absent on another — additive, always structurally valid.

Compatibility

Fully additive and optional:

• Older documents (no transcript) stay valid.
• Refs the platform didn't generate from text simply omit it.
• Older readers ignore the unknown field (structural tolerance is a documented property of both twins).

Downstream (separate PRs, not in this one)

This is the shared-contract half. To reach the screen it needs:

1. The platform to emit transcript alongside a generated clip when it produces/freezes the flow document (source text already stored), and pick up the new @wavekat/flow-schema.
2. The voice daemon to bump wavekat-flow, read prompt.transcript() for each audio prompt, thread it into the flow-detail asset view, and render it next to the play control.

Testing

• cargo test -p wavekat-flow — 35 unit (incl. new as_text/transcript/YAML round-trip in model_ext) + 3 conformance, green.
• pnpm --filter @wavekat/flow-schema test — 86 pass (new prompt.test.ts; conformance corpus picks up the new fixture); typecheck clean.
• cargo fmt --all --check + cargo clippy clean.
• Re-running pnpm gen produces no drift.

🤖 Generated with Claude Code