feat(observability): structured termination reason for every run

[ZuyiZhou]

Problem

When a run ends abnormally, the result payload carries no machine-readable cause. Operators are left doing forensic fingerprinting — e.g. inferring "launch stall" from n_messages == 0 && elapsed ≈ watchdog timeout, or diffing timestamps to separate "gateway rejected the request" from "agent deadlocked internally" from "model stopped making progress".

In batch usage this is the single largest operational cost: for one incident class, roughly half of the debugging time went into merely classifying failures before any actual fixing could start.

Proposal

Every terminated run should emit a structured termination record in the result artifact itself (not only in logs), e.g.:

{
  "finish_reason": "llm_call_timeout",
  "detail": { "last_error": "...", "http_status": 429, "turn": 17 }
}

Suggested enum, minimally:

• completed — normal finish
• max_turns / budget_exhausted
• context_limit_salvaged — see bug: context-limit salvage silently promotes think fragments into final_answer; think-tag misalignment on backends without a reasoning parser #152
• llm_call_timeout vs backend_error(code) — these two being distinguishable is the core ask
• no_progress — loop-break / degeneration interventions
• aborted — external kill

Design notes:

• The record belongs on the trajectory/result payload so downstream consumers can filter programmatically.
• Pairs naturally with a heartbeat/progress event stream (feat(observability): heartbeat / progress events for long-running tasks #153): heartbeat covers the run's lifetime, termination reason covers its end.