feat: per-scene sfx via elevenlabs (milestone a of #25)

[cuio]

First audio-bed retention lever from the SFX/Music + Gemini plan (#25). Per-scene generated sound effects landing on the existing SFX lane that PR #20's placeholders wired.

Summary

	What
Backend	generateSoundEffect() ElevenLabs client · SFX manifest (append-only, three anchors) · 4 new storyline routes (suggest / generate / get / delete) · assembler emits <audio data-track-index="3"> per manifest entry
Frontend	New 🔊 Add SFX per-card action · two new stacks per card (amber "proposals" + emerald "SFX on lane") · HTML5 audio audition + Remove on each applied entry
Tests	10 new anchor-math cases — scene-start / scene-end (with SFX-longer-than-scene clamp) / accent-word (mid / last / past-last / negative / empty narration fallback)

How it composes

[🔊 Add SFX] click on a card
        │
        ▼
POST /storyline/sfx-suggest   ← Haiku 4.5, ~$0.001
        │
        ▼ returns 1-3 { prompt, durationS, anchor, rationale }
[Amber proposal stack on the card]
        │ user clicks 🔊 Generate
        ▼
POST /storyline/sfx-generate  ← ElevenLabs Sound Generation, 1 credit
        │
        ▼ writes assets/sfx/<sceneId>-<id>.mp3, appends manifest
[Emerald "SFX on lane" stack — HTML5 audio audition]
        │ next assemble reads the manifest
        ▼
<audio data-track-index="3" data-timeline-group="sfx">  on the SFX lane

Architecture notes

1. Manifest, not script.json fields. SFX entries live in assets/sfx/sfx.manifest.json. The assembler reads it at assemble time. Keeps script.json purely about narration + scene visual decisions.
2. Three anchors, no free-form offsets: scene-start, accent-word, scene-end. Covers ~95% of cinematic usage. Resist adding offsets until users ask.
3. Soft delete only. DELETE /sfx/:entryId removes the manifest entry but leaves the audio file on disk so the user can recover by hand-editing JSON.
4. Suggest and generate are separate routes. Lets the user scan multiple Haiku ideas before paying ElevenLabs credits to generate one.
5. Cost-tracked end-to-end. Suggest logs script.storyline.sfx.suggest (Haiku); generate logs script.storyline.sfx.generate (ElevenLabs). The CostBadge already aggregates by op.

Test plan

• 729 core tests pass (was 719; +10 anchor-math)
• 281 studio tests pass
• Lint, format, typecheck clean
• Live verify: Storyline tab loads 16 scene cards, each has the 🔊 Add SFX button, zero console errors
• Manual: click 🔊 Add SFX on s04 (the long-narration hook). Confirm 1-3 amber proposals appear with prompt/duration/anchor/rationale.
• Manual: click 🔊 Generate on one. Confirm the file lands at assets/sfx/s04-sfx-….mp3, an emerald "SFX on lane" row appears with playable audio, and the master HTML's SFX lane shows the new clip on the timeline.
• Manual: Remove the entry. Confirm the manifest entry is gone but the file stays on disk.

Milestone A of #25 — what's next

	When	What
B Music	next	Multi-scene background music tracks via Eleven v3 Music polled-job pattern
C Gemini render review	after B	Analyse the rendered MP4 → structured retention feedback + scroll-risk windows
D Image analysis	after C	Auto-detect role / vibe / suggested treatment on upload
E Per-scene scroll test	after D	"Would they scroll?" frame-sample prediction with one-change fix proposal
F Retention map	last	Horizontal strip at the top of the Storyline tab — green/amber/red squares per scene

Each subsequent milestone reuses the manifest pattern + applyPatch pipeline that this PR puts in place.

🤖 Generated with Claude Code