Nice-to-haves

What's not shipped. Single source of truth — when you finish one, delete it from this file.

The earlier improvement-plan.md (4-round survey, 15-item ranked plan) is closed: every item shipped, see git log --oneline --grep="item #". The list below is what surfaced after, plus a few the original plan deferred on purpose.

Ranked by leverage

1. Scene-only audio re-synth in scene_edit_mode

Effort: 2–3 hr · Impact: ★★ · Risk: low

Today, when scene-edit-mode regenerates one scene's transcript, the whole episode still re-synthesizes audio (TTS chunks need to flow continuously). But only the chunks whose transcript changed need to be re-rendered — the others could be cached by transcript hash and reused as-is. Cuts iter-2 audio cost by ~80% on top of the scene-edit-mode trigger condition. Keys on the same content hash idea Lyria/ElevenLabs already use. (Partial: rerender_chunks exists; the cache-by-transcript-hash optimisation is the remaining piece.)

2. New judge axes for top-tier discrimination

Effort: 1 day + recalibration · Impact: ★★ (only if saturation persists) · Risk: low

Once the citation cap (item 1 of original plan) and anchored rubric (item 8) ship — both done — the rater might still saturate above 9.5. If it does, add 2–3 judge axes that are harder to max: re_listenability, surprise_at_specificity, cultural_anchoring. Reserved for after a measurement shows saturation actually persists; don't add prematurely.