Bound PR-poller GitHub GraphQL budget: dormancy-tiered cadence + kill-switch

[anutron]

Why

The daemon PR-status poller re-queries every eligible branch every 60s. On a real working set (208 tasks, 82 non-terminal branches) that is ~82 GraphQL branch-lookups/minute ≈ ~4,900 GitHub GraphQL points/hour — essentially the entire 5,000/hr budget, leaving the GraphQL bucket exhausted for everything else.

The earlier fixes (#773 terminal-state skip, #782 per-repo batching) helped but did not solve it: GitHub's GraphQL budget is cost-based (~1 unit per branch resolved), not request-based, so collapsing 82 lookups into one HTTP request still costs ~82 units. Measured live with both fixes deployed: ~55–97 graphql/min, on pace to drain the budget every hour. 78 of the 82 eligible branches had no PR at all (none, non-terminal so never skipped) and 62 had no activity in over a week — ~95% of the budget was spent re-confirming that dormant branches still have no PR.

What changes

• Dormancy-tiered cadence in pollPRStatesOnce. prPollCadenceStride derives a per-task stride from the most recent lifecycle timestamp (max(ended_at, started_at, created_at)): within 1h → every cycle, 1h–24h → every 5th, 24h–7d → every 15th, >7d → every 30th.
• Open-PR hot floor: a branch whose cached state is an open PR (draft/awaiting-review/changes-requested/approved) is polled every cycle regardless of age, so an externally-merged/approved PR still surfaces within ~60s.
• Spread by fnv(task-id) (prCadenceSelects) so each tier's tasks distribute across the stride window instead of a thundering herd on one cycle.
• Operator kill-switch: a pr-poller.disabled sentinel file under the data dir pauses the poller (zero GitHub queries) with no daemon restart; remove it to resume.

Expected steady-state cost on the current working set: ~10 branch-lookups/cycle (≈600 pts/hr) instead of ~82 (≈4,900/hr) — and it improves as tasks idle. Cadence-deferred tasks count toward skipped, preserving the len(eligible) == written + errored cycle invariant. pollCycle is in-memory (re-phases harmlessly on a daemon bounce). No schema change — cadence is derived from existing tasks timestamps and the existing task_meta pr cache.

Testing

• New table tests for prPollCadenceStride (every tier + open-PR floor + most-recent-ts wins), prCadenceSelects (stride-1 always, stride-30 once-per-window), and spread-across-cycles.
• Integration tests: a dormant unselected task issues no GraphQL query and keeps its cache; an open-PR/old task is still polled; kill-switch pause + resume.
• make pre-pr green: build, vet, fmt-check, lint-pr, test-cover-gate (89.6% ≥ 88 floor). (vuln reports only pre-existing stdlib advisories, which CI runs continue-on-error.)

Specs: OpenSpec change throttle-dormant-pr-poll (delta on the pr-status capability).

🤖 Generated with Claude Code

Co-Authored-By: Claude noreply@anthropic.com

[github-actions]

Merging this branch will increase overall coverage

Impacted Packages	Coverage Δ	🤖
github.com/drn/argus/internal/daemon	85.52% (+0.37%)	👍

---

Coverage by file

Changed files (no unit tests)

Changed File	Coverage Δ	Total	Covered	Missed	🤖
github.com/drn/argus/internal/daemon/daemon.go	84.81% (+0.85%)	428 (+29)	363 (+28)	65 (+1)	👍

Please note that the "Total", "Covered", and "Missed" counts above refer to code statements instead of lines of code. The value in brackets refers to the test coverage of that file in the old version of the code.

Changed unit test files

• github.com/drn/argus/internal/daemon/daemon_test.go
• github.com/drn/argus/internal/daemon/pr_poll_test.go