bench a provider when the provider is what failed a CoS agent run

[atomantic]

Summary

agy TUI agents were failing in bursts with the Antigravity banner "We're finishing verifying your account eligibility. This usually takes a moment. Please try again shortly."

Detection for that banner already existed (shipped 2026-07-27) and correctly failed the run in ~3s instead of waiting out the 180s idle reaper. Two things downstream of it were wrong:

1. The banner was classified actionable, so resolveFailedTaskDecision blocked the task outright — over a condition whose own message says it clears itself.
2. Nothing benched the provider. agentFinalization only marked a provider unavailable for usage-limit/rate-limit, so Antigravity stayed "available" and the next dequeued task picked it again and died identically. A queue of tasks could burn down in under a minute with every task parked.

Confirmed against two real runs on this install (agent-70d755ef, agent-ca367708): both success: false, category: auth-error, requiresFallback: true, ~3s duration, provider never benched.

Changes

• server/lib/providerCooldown.js (new) — the cooldown policy (COOLDOWN_MS_BY_CATEGORY, the request-specific never-bench rule, resolveProviderBench) extracted out of promptRunner.js, which was the only owner. The prompt cascade and the agent path previously disagreed about the same input: auth-error benched 15m via the prompt path and never via the agent path.
• agentFinalization.js — benches on any provider-origin failure, resolving the window through the shared policy. The gate is origin === 'provider' (the provenance flag from Require provider/runner provenance before excluding environmental failures from task learning #2642), so a loose keyword match on a repainted TUI transcript — which is a whole session of text the agent itself wrote — can never bench a provider.
• errorDetection.js — the eligibility signal is actionable: false (task retries instead of blocking) and origin: 'provider' (it is provider chrome). Signals stay actionable-by-default; only one that says it self-clears opts out. No benchMs on the payload — the vendored toolkit stays a classifier and the host owns remediation.
• finalizeHelpers.js — hoisted the immediateFallbackAnalysis short-circuit above a ~16KB spool read whose result was discarded. This path fires more often now.

Behavior preserved: markUsageLimit keeps its dedicated marker (it parses its own "resets 5pm" window); request-specific failures (bad model id, content refusal, off-shape response) still never bench.

Test plan

• npm test --prefix server — 1245 files / 25664 tests pass.
• New server/lib/providerCooldown.test.js covers the policy directly (skip/usage-limit/per-category/default/missing-category).
• New server/services/agentFinalization.providerBench.test.js drives finalizeAgent and asserts what gets benched. It imports the real detector and the real cooldown table so the assertions can't drift from what ships.
• Detection verified against the actual ANSI-stripped bytes from a real failing agy transcript on this install (the banner renders as two lines with a leading space from the erase-to-start sequence — the fixture in errorDetection.test.js is that shape, not idealized text).
• The provenance gate is pinned by a bypass probe: planting the old category-based predicate back makes the two output-scan cases fail, then they pass again once restored.

Review

Reviewed by agy, which caught that the first cut kept a || category === 'usage-limit' || category === 'rate-limit' fallback next to the new provenance check — preserving old behavior but contradicting the rule the code stated, and leaving a healthy provider benchable off an agent's own test output printing "rate limit". Fixed in d8c2f40, with the test gap that hid it closed.