Handle Cursor resource exhaustion retries

[andrei-hasna]

Summary

• Add a narrow provider_capacity health classification for Cursor resource_exhausted failures when stderr includes a cursor.sh host.
• Treat provider_capacity as provider-gated for scheduler retry backoff and retry-pending health warnings while attempts remain.
• Keep exhausted capacity failures routeable/high-priority and guard against broadening auth/rate-limit pending retries into provider warning health.

Validation

• bun test src/lib/health.test.ts src/lib/scheduler.test.ts src/lib/workflow-runner.test.ts (post-amend: 71 pass, 0 fail)
• bun run typecheck
• bun run build
• bun run test:boundary
• git diff --check
• git diff --cached | todos redaction scan (no secret-like values detected before commit/amend)
• bun test passed before the final self-review narrowing change (671 pass, 15 skipped live Postgres tests); final changed surface was rerun with the focused suites above.

Task

• Todos task: 4ee2501e-75ec-4438-ac43-80e5db1e9760
• Commit: d1a819764fb83e7fe05b3a94725af8a591d17069