Audio-clip reuse: central content-addressed bake cache (script_hash+voice) — bake-once across lessons

[gyasis]

The biggest single TTS saving. The bake (skill/lib/audio-bake.ts) already content-addresses each clip — script_hash = sha256(JSON(segments)) (hashSegments) — and reuses it WITHIN a lesson (audio_clips table, status reused). Lift that reuse to the central catalog so it's CROSS-lesson.

Design

• Central content-addressed audio store keyed by (script_hash, voice) → mp3.
• Before synthesizing ANY segment in ANY lesson: HIT → copy/link existing mp3, NO TTS call; MISS → bake once, store, reuse forever.
• A class-owned foundation narration bakes ONCE and is reused by every member/disease lesson that inherits the same text + voice. Pairs with Catalog: make a Foundations/exposition block a reusable, class-owned stable unit #9 (foundations as reusable unit) + Class→disease inheritance: disease deep-dive PULLs the class Foundations block #11 (class→disease inheritance) + Reuse policy: REPEAT / PULL / CITE / HYDRATE decision at generation time #10 (PULL).
• Economy-first ($5K-scar): hash-dedup BEFORE the paid synth call. Voice is part of the key (Lucrezia ≠ pauls_tutor).

Caveat / dependency

Reuse needs byte-identical script text + same persona → lessons must INHERIT the same foundation text, not paraphrase. (This is the concrete reason inheritance must share TEXT.)

Evidence (2026-06-24)

Phase-2 (acute deep-dive) and Phase-3 (pericardial-disease class survey) each re-baked the same pericardial foundation narration — pure waste a (script_hash, voice) cache eliminates.

PRD: prd/chiron_lesson_scope_and_reuse_2026-06-24.md §4b · Epic: #6