subagent cost store: decay stale samples or probe-admit when a learned p90 outlives its roster

[bolichen97]

Deferred from the Design Review Watch on #13294 (span=b2211c56f0c9).

Owner: @bolichen97
Due: 2026-10-24

Task

The spawn guard now prices every unmeasured dedicated start at the learned per-run p90 (never below agent.subagent_cost_gb). The p90 is only ever refreshed from runs that actually execute, so on a host whose real per-run cost DROPPED (a heavy MCP roster removed, a backend switched) and whose free memory never clears spawn_min_memory_gb + old p90, deferred starts produce no new samples and the stale figure never self-corrects. #13294 makes that state visible (the deferral names the price, its source and the reset: delete subagents/cost_samples.jsonl under the data home, or lower spawn_min_memory_gb) but leaves the remedy manual.

Pick one and pin it with a test:

• Sample decay: age out samples older than N days (or beyond a wall-clock window) in compact_cost_log / read_learned_cost, so a roster change is forgotten without operator action.
• Probe admission: when zero dedicated workers are live and the ONLY thing blocking a start is the learned reserve (the configured-cost reserve would admit it), admit exactly one start so a fresh sample lands; its observed RSS then re-prices the next one.

Not a goal: weakening the reserve while dedicated workers are live -- that is the window #13294 exists to close.

Pointers: _startup_memory_reserve_gb / _startup_cost_gb in src/kiro_crew/subagent.py, _refresh_learned_cost_impl in src/kiro_crew/subagent_manager/monitoring.py, read_learned_cost in src/kiro_crew/subagent_cost.py.

[github-actions]

Automated triage summary — generated from the issue text alone, before any maintainer has read it. Treat every line as a starting point, not a verdict.

Follow-up task deferred from the #13294 design review: the spawn guard's learned per-run p90 cost for dedicated subagents only updates from runs that actually execute, so when a host's real cost drops (MCP roster removed, backend switched) and memory never clears the stale reserve, starts stay deferred and the figure never self-corrects. Asks maintainers to choose between aging out old samples or admitting one probe start, and pin it with a test.

_{Posted by issue-summary.yml. The model had no access to the repository — it read only this issue and the titles of recent ones.}

[bolichen97]

Progress from #13294: a 30-day sample horizon (_SAMPLE_MAX_AGE_SECS) landed there, so a price learned under a removed workload expires on its own within a month, and a bucket without dedicated history is priced from the configured cost rather than the heaviest known bucket. What remains for this issue: a faster decay, or probe admission, for hosts that cannot wait a month.