fix(flow-schema): decode audio prompt transcript

[wavekat-eason]

Why

The platform bakes a transcript into every audio prompt when it publishes a flow ({ audio: vprompt_…, transcript: "…" }). On re-parse, the shared schema logged a warning for each such node:

node "welcome" prompt: unknown field "transcript" is ignored by the engine

Root cause

#30 made transcript a first-class optional field on the audio-prompt model — it regenerated the TS model and bundled JSON schema and added promptTranscript()/Prompt::transcript(). But the hand-written safe-subset YAML parser in packages/flow-schema/src/parse.ts was never updated. Its decodePrompt still knew only ['audio'], so it:

1. warned that transcript was an unknown field, and
2. dropped it — promptTranscript() returned null for any audio prompt decoded via parseFlow, contradicting what feat: carry an audio prompt's transcript text #30 shipped.

Fix

Teach decodePrompt the field:

• add transcript to the known-field list (no more spurious unknown_field warning), and
• carry it onto the returned prompt, validated as a string (a non-string transcript is now an expected_string error, consistent with the model's transcript?: string).

The Rust twin already carries it (serde silently ignores unknowns and the Audio struct has the field), so this closes the TS-only gap and brings both twins back in sync.

Tests

• New parse test: an audio prompt with transcript round-trips through decode, promptTranscript() returns the words, and zero unknown_field warnings are emitted.
• New negative test: a non-string transcript yields expected_string.
• Full flow-schema suite (88 tests) + tsc --noEmit green.

🤖 Generated with Claude Code