refactor: scheduled_testing_eligible column + derivation bridge (PR A)

[petterlindstrom79]

Summary

• Adds test_suites.scheduled_testing_eligible BOOLEAN NOT NULL DEFAULT FALSE and a Drizzle migration that backfills eligible = TRUE where external_cost_cents = 0 (preserves current scheduler behavior exactly).
• Swaps the 3 scheduling readers in test-scheduler.ts (dispatch + 2 observability) and 1 diagnostic-script reader (investigate-singapore.ts) to filter on the new column. external_cost_cents returns to being billing-only.
• Adds startup block 0066_reconcileEligibilityFromCost as an interim derivation bridge — re-derives eligibility from cost at every boot so any row inserted by the 12 untouched INSERT call sites lands at the correct value.

Why now

The May 2026 Haiku token spike (PRs #84/#85/#86/#87) was structurally possible because external_cost_cents = 0 did double duty as billing data and scheduling signal. PR #46's cadence flip (24h → 1h) plus PR #49's deferred Anthropic-Haiku cost-bump silently turned billing-data lag into a scheduling regression for 73 LLM caps for 7 days. PR #85 shipped a CI assertion as the interim guard. This PR decouples the underlying coupling so the assertion is defense in depth, not load-bearing.

Scope split

PR A (this PR):

• Schema + migration + dispatch swap + observability + script + startup block 0066 + Notion writes.
• INSERT call sites and CI assertion are not touched.

PR B (separately tracked, Notion To-do):

• Force scheduledTestingEligible: boolean explicit at all 12 INSERT call sites.
• Remove block 0066.
• Rewrite CI assertion to check LLM-importer + ALWAYS_LLM_CAPABILITY_COSTS + eligible = FALSE invariant directly.

Notion

• New DEC: DEC-20260511-B — Decouple scheduled_testing_eligible from external_cost_cents (PR A). Refines DEC-20260503-B, does not supersede.
• Existing structural-follow-up To-do 35d67c87082c81148ee4fc88f1671776 moved to In progress with status note.
• Capability onboarding pipeline page updated with interim-derivation note.

Verification

• Type-check: pre-existing routes/mcp.ts strale-mcp/tools errors unchanged; no new errors.
• Vitest: 540 passing / 11 skipped / 1 failing (app.classify-error.test.ts is the pre-existing failure noted in the prompt; unrelated to this PR).
• Updated startup-migrations.test.ts BLOCKS-list assertion (8 → 9 blocks; new block appended in order).
• Grep confirms no remaining external_cost_cents = 0 / > 0 scheduling readers in src/ outside the legitimate billing-data UPDATE blocks (0062–0065) and block 0066's intentional derivation.

Closing-steps rule walk

• Rule 4 (source-health integrity): dispatch swap is the carve-out PR fix: bump external_cost_cents on always-llm capabilities to gate hourly testing #85 named.
• Rule 7 (bug-fix framework): not invoked. Structural prevention, not a bug fix.
• Rules 8 + 9: hand-written Drizzle migration + same-commit schema.ts sync, per DEC-20260420-A.
• Rule 11 (DEC supersession): does NOT supersede DEC-20260503-B. Refines. No sweep needed.
• Rule 12 (audit follow-up tests): dispatch / observability / script swaps covered by backfill parity (same cap set pre/post); block 0066 covered by its post-condition DO block.

After-deploy plan

cd apps/api && npx drizzle-kit migrate
psql $DATABASE_URL -c "\d test_suites" | grep scheduled_testing_eligible
psql $DATABASE_URL -c "SELECT COUNT(*) FROM test_suites WHERE scheduled_testing_eligible = TRUE"
psql $DATABASE_URL -c "SELECT COUNT(*) FROM test_suites WHERE external_cost_cents = 0"

Both counts must be equal. Then monitor the next hourly dispatch tick — cap set should match the pre-deploy set exactly.

Test plan

• Drizzle migration runs locally with the post-condition DO block passing
• Backfill row counts match pre-migration set
• All 3 swapped scheduler queries return identical row sets to the pre-swap version
• Startup-migrations BLOCKS-list test passes with new 9-block expectation
• Full vitest run: 540 passing
• After Railway deploy: \d test_suites shows new column; both WHERE eligible = TRUE and WHERE cost = 0 counts equal
• After Railway deploy: next hourly dispatch logs same cap set as pre-deploy

🤖 Generated with Claude Code