fix(store): skip completed usage totals backfill

[Gajesh2007]

Summary

Make the usage_totals backfill genuinely one-time and lossless under concurrent writes. Existing deployments adopt the already-populated counter without scanning usage; fresh databases use a durable two-phase cutoff so the historical scan runs without holding a writer-blocking table lock.

Linked issue

N/A

Before

flowchart LR
  subgraph Behavior
    B1[Coordinator startup] --> B2[Aggregate every usage row]
    B2 --> B3[Discover usage_totals conflict]
    B3 --> B4[Bind HTTP listener late]
  end
  subgraph Code
    C1[NewPostgres] --> C2[migrate statement loop]
    C2 --> C3[INSERT SELECT FROM usage ON CONFLICT]
  end

Loading

After

flowchart LR
  subgraph Behavior
    B1[Coordinator startup] --> B2{Applied marker or existing counter?}
    B2 -->|Yes| B3[Skip historical usage entirely]
    B2 -->|No| B4[Briefly fence inserts and persist max usage ID]
    B4 --> B5[Resume writers against zero counter]
    B5 --> B6[Aggregate only historical IDs]
    B6 --> B7[Atomically add baseline and mark complete]
    B3 --> B8[Continue startup]
    B7 --> B8
  end
  subgraph Code
    C1[NewPostgres] --> C2[migrateUsageTotals]
    C2 --> C3[prepareUsageTotalsMigration]
    C3 --> C4[usage_totals_backfill_state cutoff]
    C4 --> C5[finalizeUsageTotalsMigration]
    C5 --> C6[schema_migrations marker]
  end

Loading

Test plan

• DATABASE_URL=... go test ./store -run 'Test(MigrateNoUngatedUsageTotalsAggregation|UsageTotalsMigration)' -count=10
• DATABASE_URL=... go test ./store -count=1
• go test -count=1 ./...
• go vet ./store

Components touched

• coordinator (Go)
• provider (Rust, legacy)
• provider-swift (Swift CLI)
• console-ui (Next.js)
• enclave (Swift)
• infra / CI / release
• docs

Protocol / interface changes

• No protocol/interface changes
• Yes — described above and matching side updated

Notes for reviewers

The old ON CONFLICT DO NOTHING was not a fast path: PostgreSQL evaluated the aggregate source before checking the destination conflict. The new checkpoint captures a stable ID cutoff while inserts are fenced only for a primary-key MAX(id), counter insert, and checkpoint insert. The full scan happens after writers resume; post-cutoff writes increment the live counter while finalization atomically adds only id <= cutoff. An interrupted process resumes from the checkpoint without double-counting.

  

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
d-inference	Ready	Preview	Aug 22, 2026 8:00am
d-inference-console-ui-dev	Ready	Preview	Aug 22, 2026 8:00am
d-inference-landing	Ready	Preview	Aug 22, 2026 8:00am