[Bl0ck154]

Summary

Implement all four requested Antigravity/Gemini 3.8 resilience improvements together against the fork's production branch:

1. Distinguish confirmed pre-response/header 504 timeouts from transient 502/503/socket resets. Prevent same-account 504 retry, release and rotate to another eligible account, preserve explicit pin semantics and request limits.
2. Optional adaptive account + model health, bounded rolling sample, 499 ignored, quarantine with cooldown/backoff, single half-open real-request probe, two-success recovery, concurrency/generation fence; state in management connections API/dashboard.
3. Optional sanitized timing telemetry for idle time, inflight concurrency, time to headers, first substantive SSE output and completed phases, with DNS/TCP/TLS null when not measurable. No synthetic paid probe traffic.
4. Hot DB-persisted Antigravity preferred failover timeouts (defaults 5000ms / Gemini 3.8 12000ms), admin-only resilience API, validation 500–120000ms, UI Settings → Resilience, revision-based write protection and per-request config snapshots.

Evidence

Production current release 3.8.52-custom.32 (unchanged during development), production fork based on SHA 44fdae6b. Live SQLite logs in the last 24h showed about 2,824 Gemini3.8 call entries: 261 504, 576 499, 1,874 200 (window-dependent). Roughly 100 same-connection consecutive 504→504 sequences by correlation ID; owner reported 113 in another window. 504 durations near 12,025ms match current Gemini3.8 start timeout. Idle-gap evidence is correlational, not root-cause proof. Original retry classifier included 504. Client 499 and 429 are not counted as model timeout health failures.

Verification gates

Unit tests: tests/unit/antigravity-adaptive-health.test.ts, antigravity-timeout-retry.test.ts, antigravity-timing.test.ts.
Integration: tests/integration/antigravity-resilience-routing.test.ts, antigravity-resilience-settings.test.ts.
All five added to selected-upstream-regression.yml alongside existing failover tests and TypeScript typecheck. These tests have not yet been validated by CI. Do not merge or deploy until green. No build/test was run on VPS per /opt/AGENTS.md; full audit and limitations in docs/audits/2026-10-24-antigravity-resilience/README.md.

Release

After green CI and review, the operator will use GitHub Actions to build and test a packaged artifact off the production VPS and deploy once with the existing canary/rollback installer, then verify authenticated Antigravity requests and hot settings. Feature gates for adaptive quarantine and timing diagnostics default OFF for safe rollout; enable via admin API after deployment and confirm actual state.