Stop reconciliation GraphQL spending after an observed reserve hold

[Nitjsefnie]

Summary

Stop an admitted reconciliation from starting more GraphQL requests after its sponsor's observed balance falls below the reserve. A held crawl skips materialization and defers its queue row to the observed reset deadline without consuming an attempt.

Related Issues and Pull Requests

Fixes #229

Changes

• Check the existing sponsor budget immediately before GraphQL transport and after successful observations, including the final page.
• Keep a reconciliation-scoped hold latched across concurrent collectors and reset expiry; ordinary owned gateway calls outside the crawl remain available.
• Settle started runs as incomplete and return the existing budget-held result, preserving worker deferral and attempt refunds.
• Add transport, concurrency, observer-failure, and worker regression coverage while retaining all six operations and the existing reserve equality/expiry behavior.

Testing

• RED tests reproduced extra issue-page requests and final-page materialization after a low observation. Disabling both transport checks caused four targeted failures; removing only the pre-request check exposed an extra GraphQL request; clearing the latch produced twelve requests instead of six; and routing the typed hold through generic error handling returned RETRY_SCHEDULED instead of BUDGET_HELD. Each mutant was restored and its focused tests returned green.
• Focused verification passed seven files and 448 tests.
• pnpm db:migrate && pnpm test --run && pnpm lint && pnpm typecheck && pnpm build passed in implementation and again in the independent final review: 118 files and 2,712 tests passed with zero skips; lint, typecheck, and production build passed.
• GitHub CI passed both verify and actionlint on exact head 40b792fabddcb47ddac3088b76ac19d4e520f8f2.
• Output findings: three existing Edge Runtime node:crypto build warnings in untouched modules; pnpm warned that four dependency build scripts were ignored; migration emitted existing-object notices.

Follow-ups / Known Limitations

Requests already active when the hold is observed may settle and spend points; this does not guarantee an absolute global balance floor. Held retries restart collection rather than resuming a cursor. A held run uses the existing FAILED run status while its queue row is deferred with its attempt refunded.

Footer

Generated by GPT-5.6 Sol (brief, orchestration), GPT-6 Astra (discovery, implementation, testing, verification, review, scoring)