fix(admin): bulk SQL upsert in Core sync phases — replace per-row $transaction that timed out at 5s in prod

[Kneesal]

Summary

All 5 Core sync phases (languages, countries, keywords, videos, dubs) wrapped 500 sequential prisma.upsert() calls in $transaction({ timeout: 5_000 }). Confirmed during the R1 prod smoke on 2026-04-27: every page hit Transaction API error: Transaction not found, logged 5 identical errors over ~16s, and wrote zero rows. 500 upserts × ~10ms over Railway's proxy = 5s with zero margin.

This PR replaces the per-row pattern with one bulk INSERT ... ON CONFLICT DO UPDATE statement per entity per page:

• Single round-trip instead of N (~100x faster)
• Atomic without a transaction wrapper (single statement is implicit-tx)
• Not subject to the interactive-transaction timeout at all

Per-phase notes

Phase	Bulk statements per page	Notes
languages	1 (Language INSERT)	page-loop, 500/page
keywords	1 (Keyword INSERT)	single batch (no paging in Core query)
countries	2 (Continent then Country)	deduped continent set first via RETURNING; Country uses the resulting (core_id → id) map for continent_id
videos	2 (Video then VideoLocale)	MANAGER protection moved into ON CONFLICT WHERE \"video\".\"source\" != 'manager'; videos that survive the WHERE drive the flattened (videoId, locale) tuples for the locale upsert
dubs	1 (VideoDub INSERT)	same MANAGER WHERE clause; orphaned dubs (no video FK) are now logged + skipped instead of throwing — one stale FK shouldn't poison the rest of a 500-row batch

Behaviour preserved

• Per-page error handling (catch, increment stats.errors, structured log, continue paging)
• Soft-delete sweep at end of each phase (still prisma.updateMany, unchanged)
• Loop termination: still break when rawRows.length < PAGE_SIZE
• Structured log event names unchanged
• MANAGER protection guarantee unchanged — semantically equivalent, enforced by SQL WHERE instead of a JS pre-pass findUnique

Helpers

src/services/core-sync/bulk-upsert.ts:

• newRowId() — wraps @paralleldrive/cuid2's createId() to mirror Prisma's @default(cuid()) since the DB columns have no Postgres-side default (verified via 0001_init migration)
• jsonbParam(value) — `${JSON.stringify(value)}::jsonb` cast required for jsonb columns under bound-parameter INSERT (cf. PG18 cast rules in apps/admin/CLAUDE.md "Common pitfalls")

Testing

• Existing sync-languages.test.ts: 3 tests updated for the new bulk shape; +2 new (bulk-INSERT SQL invariant, bulk-error path). Switched vi.clearAllMocks → mockReset so queued mockResolvedValueOnce values don't leak across tests
• New test files for the previously-uncovered phases: sync-keywords.test.ts, sync-countries.test.ts, sync-videos.test.ts, sync-dubs.test.ts. Each covers happy path + SQL invariant (INSERT INTO, ON CONFLICT ... DO UPDATE, no \$transaction) + edge cases. Entity-specific tests:
  • countries: deduped continent INSERT + RETURNING-driven Country INSERT; skip continent INSERT when none in batch
  • videos: MANAGER protection — when RETURNING omits a coreId, no VideoLocale upsert fires for it; empty-locale-set short-circuit
  • dubs: missing-video-FK skip path

Test Files  71 passed (71)
     Tests  951 passed | 1 todo (952)   ← +24 from baseline

Lint + typecheck clean.

Post-Deploy Monitoring & Validation

• What to monitor
  • Logs: service=forge-admin event=core-sync.phase.complete (existing structured event) — successful runs should show non-zero created/updated and durationMs of seconds, not the previous 24s/5-error pattern.
  • Failure signal: any core-sync.<entity>.error event with message matching Transaction API error (the old failure mode) — should be impossible after this fix; if it appears, escalate.
• Validation checks (post-deploy, against admin prod)
  • Fire from a logged-in admin browser console:
    ```js
    const r = await fetch('/api/graphql', { method:'POST', headers:{'content-type':'application/json'}, credentials:'include', body: JSON.stringify({ query: 'mutation { triggerSync(scope: "languages", incremental: false) }' })});
    console.log(JSON.stringify(await r.json(), null, 2));
    ```
  • Expected: phases: [{ phase: \"languages\", updated: 10000+, errors: 0, durationMs: <few seconds> }] — replaces the previous errors: 5, updated: 0, durationMs: 24588 shape.
  • Then SELECT count(*) FROM language against admin prod Postgres → matches Core's language count.
  • Then triggerSync(scope: \"all\", incremental: false) to populate the full catalog.
  • Then re-fire triggerSceneEmbeddingBackfill(coreIds: [\"2_FabioStory\"], locales: [\"en\"]) — should now produce non-zero totalTargets (was 0 because admin's video table was empty).
• Failure signal / rollback trigger
  • If triggerSync(scope: \"languages\") returns errors: > 0 post-deploy, capture the structured core-sync.<entity>.error log line, revert this PR, and investigate. The fallback is to ship the simpler timeout bump (5_000 → 60_000) discussed in the PR thread.
• Validation window & owner
  • Window: ~10 min after deploy, while smoke-testing.
  • Owner: Nisal — this PR is the prerequisite for completing the R1 prod smoke that PR feat(admin): split S3 client — read manager artifacts from manager's bucket #845 unblocked the bucket-routing for.

Refs

• apps/admin/CLAUDE.md R1 runbook "R0 prereq" paragraph (added 2026-04-27 alongside PR feat(admin): split S3 client — read manager artifacts from manager's bucket #845)
• docs/handoffs/2026-04-21-admin-migration-r1-smoke-and-r2-handoff.md
• PR feat(admin): split S3 client — read manager artifacts from manager's bucket #845 — bucket-routing fix that surfaced this separate Core-sync-never-ran prerequisite

🤖 Generated with Claude Code — Claude Opus 4.7 (1M context) via Compound Engineering ce:work

[railway-app]

🚅 Deployed to the forge-pr-846 environment in forge

4 services not affected by this PR

• @forge/cms/db
• @forge/cms
• @forge/manager
• @forge/web