fix(headless): classify a benchmark trial by who ended it, not by how the exception reads

[Astro-Han]

Summary

A benchmark trial whose agent phase ended abnormally was classified by the wording of its exception. isBudgetExhaustedTrialException matched two pinned messages; only on a match did the runner read the reward / verifier outcome / cell output. Everything else hit the exitCode !== 0 gate and was thrown as infra — dropped from the Pass@1 denominator — even though both harnesses run the verifier after an agent-phase exception, so the reward was already on disk.

The root defect is narrower than "the regexes are brittle": the termination reason arrives typed and was being thrown away. Harbor records exception_type as Python's type(e).__name__ (harbor/models/trial/result.py:31), and readTrialException flattened it to "Type: message" before any caller could use it. Every downstream decision — the regexes, the exit-code gate, the settle predicate — was re-deriving that discarded fact by matching prose.

So keep it structured. Classification now reads two facts, who ended the agent phase and whether the verifier reached a verdict:

	verdict	no verdict
agent-owned (AgentTimeoutError, host-cell deadline, NonZeroAgentExitCodeError)	scored	budget_exhausted
external (Docker, harness, unrecognized)	infra (on a non-zero exit)	infra

• agent_budget — Harbor's own AgentTimeoutError class is unambiguous by type. The host cell's deadline arrives as a generic RuntimeError, so it is the one termination still matched by message; that is sound only because this repo raises it (harbor/maka_agent.py) and harbor-adapter.test.ts pins both ends together.
• agent_exit — NonZeroAgentExitCodeError is the agent's own process ending, so a verdict on the workspace it left is a real result whatever the exit code. Nothing fabricates a deadlineSettlement for it; taskCompletedScoringProjection already scores on verifierGrade !== undefined, so the cell keeps its honest status / errorClass.
• external — unrecognized types land here deliberately. The recognized budget shapes are matched by type, so nothing regresses when upstream rewords a message.

The terminal-provider-request gate had the same defect in miniature. A boolean "this trial settled" exempted every non-completing tail request. Only a client-side teardown (aborted, set from signal.aborted) is explained by the agent phase ending — interrupted and failed are the upstream's doing and now stay infra however the trial settled. This tightens the pre-existing budget path in the safe direction: a provider outage is excluded from the denominator rather than recorded as the agent's zero.

Arm symmetry. claude_code_agent.py never emitted deadlineSettlement — not a missing assignment but a missing field in _write_cell_output — so the same termination was labelled budget_exhausted for Codex and a plain runtime failure for Claude Code. Both arms now report it, pinned by extending the existing cancel_codex() behavioral seam rather than by matching adapter source text.

Infra errors now name the trial exception in their message. The WAL persists only error: errorMessage(...) and drops detail, which is why 225 infra events in the current run carried no trace of the timeouts inside them.

Refs #1970. Independent of #2047.

Verification

• npm run test -w @maka/headless — 1337 pass / 1 skip, three consecutive runs (baseline 1326 pass / 1 skip).
• npm run lint, npm run format — clean.
• Mutation battery, 12 mutations, each caught: every branch of classifyTrialTermination; both conjuncts of the settle predicate; the terminal gate exempting everything / exempting nothing; pier settling external types; and the three adapter mutations a source regex could not see (flag moved into finally, emission guard inverted, emission deleted).

Review focus

Six independent reviews (deepseek-v4-flash via pi; two rounds × design / correctness / test-quality lenses) returned no P0 or P1. The round-one convergent finding — that the terminal-provider-request gate silently undid the fix — is addressed above. Round two's strongest finding was that the arm-symmetry test was a source-text proxy when an executable seam already existed; that is now fixed, and the two semantic mutations that defeated the regex are pinned.

Known and deliberately not addressed here:

• external is narrower than it looks. Both harnesses catch only (AgentTimeoutError, NonZeroAgentExitCodeError) in _run_agent and reach _run_verifier() only from there (harbor/trial/single_step.py:75-86, trial/trial.py:286-305); any other exception routes to _recover_outputs(), which does not run the verifier. So a conclusive verdict on disk is already evidence of an agent-owned termination, and both harnesses exit 0 on a recorded agent-phase exception. The external row therefore fires mainly on the non-zero-exit path. Collapsing it — letting evidence decide scoring, telemetry decide fairness, and type decide only the label — is the smaller structure, but it is a fourth restructure on the eve of a full rerun and is not worth the risk tonight.
• Pier's NonZeroAgentExitCodeError has a subclass hierarchy (ApiRateLimitError, ContextWindowExceededError, AgentSafetyRefusalError, …) that type(e).__name__ records as leaf names, so agent_exit does not reach them. Not a regression — they behave exactly as on main — but the improvement does not reach the Pier arms. The collapse above would fix this for free.
• The aborted discriminator is event-order dependent. If a connection dies in the same instant as the agent kill, the proxy may record failed/interrupted before signal.aborted is set. The right owner is the proxy (classify teardown by client-connection state, not by which error arrives first), not the runner gate.
• harbor/maka_agent.py still raises a generic RuntimeError for the host-cell deadline, so one message regex survives. A distinct exception class would remove it, but that is a cross-language contract change and a mistake there silently reclassifies timeouts as external.