refactor(style): canonical v2 style config with hash-stable projection

[tombeckenham]

Summary

Re-implements the #858 style-config redesign against current main, replacing closed PR #863. Same design goal — one canonical StyleConfig with grouped look/motion signatures — but with the migration strategy inverted: instead of a null-sweep that erases all staleness data, hashing is shape-stable by construction, so the reshape ships with no migration at all.

v2 shape

{ version: 2, summary?, tone?,
  look:   { mood, artStyle, lighting, colorPalette, colorGrading, medium?, texture?, composition? },
  motion: { camera, shots?, pace?, energy? },     // camera <- v1 cameraWork
  references: [] }                                 // <- v1 referenceFilms

The speculative design/casting/sound/source groups from #863 are cut until their consumers (Phase 3/4) exist.

Why no migration

All hashers (visual/motion prompt hashes + computeStyleConfigHash) now share one projection, styleConfigHashBody, which keeps the legacy flat key names and joins the new optional fields only when authored. A v2 config with no authored refinements therefore hashes byte-identically to its v1 row — pinned by a test asserting deep equality with the raw v1 object. PROMPT_INPUT_HASH_VERSION stays 4; no columns are nulled; genuinely-stale artifacts keep their banners. Authoring shots/pace/energy later flips staleness — correctly, because that IS an input change.

What changed

• src/lib/style/style-config.ts (new): v2 schema with version: 2 tag, validated migrateStyleConfigV1ToV2 (output is schema-parsed — a constraint-violating legacy blob fails loudly instead of flowing into prompts), parseStyleConfig, StyleProjection + toStyleProjection, styleConfigHashBody, and total (never-throwing) UI accessors.
• Honest column typing: styles.config is StoredStyleConfig (v1 | v2) until backfill — direct .look/.mood access does not compile; every read boundary goes through parseStyleConfig (prompt-context, launchers, sequence-characters/locations, style catalog, enhancer, scripts/analyze-script.ts).
• Enhancer: EnhanceStyle.config is whole-or-absent parsed v2 (no more Partial), tags always string[]; toEnhanceInputs up-converts at the row boundary so UI and API feed identical shapes; prompt renders the motion refinements when authored.
• UI: gradient/tiles/detail dialog use total accessors — a malformed row renders empty instead of a ZodError blanking the grid; shot-staleness catches escalate ZodError (data corruption) to logger.error while transient failures stay warns.
• Optionality: category required on create (via .extend — drizzle-zod keeps nullable columns optional under refinement callbacks), tags/useCases default [], references defaults [], config whole-or-omitted on update. No existing caller sends creates without category (the create-style UI doesn't exist yet).
• Templates: all 81 stay v1-authored in one file; the seed mapper converts through the shared validated converter; a test loops every template through it.
• Backfill: scripts/backfill-style-config-v2.ts validates v2-shaped rows instead of skipping them blind (corrupt "v2-shaped" blobs are exactly what it exists to find), converts+validates everything before the first write, skips updatedAt, and is idempotent.
• e2e fixtures: recorded match.userMessage strings truncated at the <DIRECTOR_STYLE>/style-block boundary so replays survive the injected-JSON reshape (responses/timings untouched).

Verification

• bun typecheck 0 errors · bun run test 2347 passed · bun lint 0 errors · bun dead-code clean · bun format:check clean
• bun db:migrate:local && bun db:seed:local — 81 templates re-sync as valid v2
• Backfill proven on local D1: migrates a real inserted v1 row, re-run reports 0, dry-run writes nothing
• bun test:e2e sequence-flow — 4 passed (script-enhance → analyze → visual/motion prompt fixtures replay through the truncated matches)

Deploy notes

No DB migration. Deploy order is free: readers tolerate v1+v2 before and after the backfill. Run bun scripts/backfill-style-config-v2.ts --d1 any time after deploy; the up-converting v1 style config at read time warn (PostHog logs, service openstory-prd) going quiet is the signal the v1 branch + StyleConfigV1 can be deleted.

Follow-ups

• Eval experiment before authoring motion-rich templates: the "richer motion signatures → better output" claim is untested — A/B a handful of templates with authored shots/pace/energy through score-style-previews / eval-style-sample-videos / compare-eval-scores before investing in a new template generation (see Style schema design: aesthetics-blob vs identity-columns split + inconsistent optionality #858 discussion; the preferred vehicle is a new v2-authored template set with the old ones hidden).
• Drop the v1 branch + retire StoredStyleConfig to plain StyleConfig once prod backfill is confirmed quiet.

Closes #858

🤖 Generated with Claude Code

[github-actions]

Cloudflare Preview deployed successfully

URL	https://pr-1174.openstory.workers.dev
Worker	pr-1174
D1 Database	openstory-pr-1174
Commit	f58cc83