feat(usage): Antigravity quota grouping and auto-ping scheduler

[mahdiwafy]

Summary

Refactor Antigravity usage parsing to group models into quotas buckets (Gemini vs Claude/GPT) as enforced by the upstream API, and add an Auto-Ping scheduler to keep windows warm.

Changes

1. Quota Grouping: open-sse/services/usage/google.js no longer loops through 10 repeated models. It now parses the response and outputs two specific buckets:

   • gemini_bucket: Represents the shared limits for all Gemini models.
   • claude_gpt_bucket: Represents the shared limits for Claude and GPT-OSS models.
   • This directly fixes the UI on /dashboard/usage to show 2 progress bars instead of 10 repeated/misleading rows.

2. Auto Ping Scheduler: Added antigravity to the Auto Ping scheduler (src/shared/services/quotaAutoPing.js & config.js).

   • Dashboard users can now toggle a lightning bolt ⚡ to auto-ping their Antigravity connection when the 5-hour window resets.
   • The ping sends a tiny 1-token "hi" request using gemini-3.5-flash-low to immediately restart the quota window.

Verification

• npm run lint:undef — clean
• Rebuilt Docker image successfully.
• Verified dashboard now shows exactly 2 buckets (Gemini & Claude/GPT) with correct remaining percentages.

[Vanszs]

Request for revision with concrete behavior requirements.

Antigravity quota grouping and auto-ping are relevant, but auto-ping changes real provider quota consumption and must not be enabled implicitly.

Please add:

• an explicit opt-in setting with a safe default off;
• provider-response fixtures proving Gemini/Claude/GPT bucket classification;
• a clear estimated marker for weekly usage calculations;
• executor-compatible ping payload tests, including proxy and account selection;
• cooldown, timeout, retry, and duplicate-scheduler protection;
• tests proving failed/aborted pings do not corrupt quota state;
• documentation of the quota cost and reset assumptions.

Please preserve existing usage history, token tracking, account fallback, proxy behavior, and RTK/Caveman/Ponytail request flow. The UI grouping can be reviewed separately from the scheduler.

[mahdiwafy]

Per your #82 review notes:

1. Explicit opt-in, default off — the scheduler only pings connections explicitly enabled via settings[antigravityAutoPing].connections[connId] === true; an empty map means no pings at all.
2. Bucket classification fixtures — added tests/unit/antigravity-weekly-bucket.test.js: gemini-* → gemini_weekly, claude-*/gpt-* → claude_gpt_weekly, error/failed rows skipped.
3. estimated marker — weekly usage now returns estimated: true (7-day request history scaled to a full week × conservative 4x multiplier), clearly marked as estimate, never presented as provider-reported.
4. Ping payload + proxy/account selection — pings go through the same executor path with resolved proxy options and refreshed credentials (existing code).
5. Cooldown/timeout/retry/duplicate protection — failure cooldown cache, duplicate-reset-key guard, min ping interval, refresh fail-soft — now unit-tested (quota-autoping.test.js): default-off, per-connection enable, duplicate skip, failed ping doesn't write lastPingedResetAt, refresh failure fail-soft.
6. Quota cost/reset assumptions documented in the tracker header (requests not tokens; weekly pool ≈ 4× the 5H pool).

All 8 new tests pass.

[Vanszs]

Request changes

Cross-check against the latest main found runtime blockers despite green build checks:

1. Antigravity is not registered in QUOTA_AUTOPING_CONFIG.providers. The scheduler iterates Object.entries(C.providers), so the new handler is never scheduled. Add the provider config with explicit per-connection opt-in and default-off behavior.
2. sendAntigravityPing() sends an incompatible executor payload. It passes messages at the top level instead of the executor's expected body shape. Add a test against the actual executor call, not a mocked sendPing handler.
3. Successful ping returns undefined after draining the stream. The caller treats falsy as failure, so lastPingedResetAt/lastPingedResetKey are never persisted.
4. No timeout/abort for reader.read(). A hanging upstream stream can hold state.running indefinitely.
5. Proxy arguments are incorrect for getAntigravityUsage(): proxyOptions is passed as providerSpecificData, leaving the proxy argument unset.
6. Weekly tracker is orphaned: no runtime usage/UI path calls getWeeklyUsage().
7. Weekly extrapolation uses first-use timestamps between buckets and can produce absurd scaling for short observations. Bound or defer estimation and test the edge cases.

Please rebase onto current main, add executor/proxy/timeout/error-path tests, and rerun targeted tests plus the full build/test checks.