feat(daemon): split opaque execution_failed into close-reason sub-details

[lefarcen]

Why

I'm taking over the run-reliability lane (#3408). Profiling current-version (0.10.0+) failures in PostHog, the single largest failure_detail is the opaque execution_failed — ~4.5k/week across providers (opencode, codex_cli, claude_code, gemini, …). It's the AGENT_EXECUTION_FAILED catch-all that fires when no text pattern matched, so on the dashboard it reads as "the agent failed and we don't know why."

But the daemon does already know more: the runtime_close diagnostic carries rpc_close_reason, which splits this bucket into three genuinely distinct shapes — mid-stream agent error (stream_error, ~51%), bare non-zero exit (exit_nonzero, ~44%), and ACP fatal close (fatal_rpc_error). That split only lived on PostHog's rpc_close_reason field, so the canonical failure_detail — and therefore the Langfuse sink, which classifies through the same classifyRunFailure — stayed opaque, and the two sinks disagreed per run.

What users will see

No user-facing UI change. This is internal failure-analytics: dashboards and Langfuse traces keyed on failure_detail now see stream_error / exit_nonzero / fatal_rpc_error instead of one undifferentiated execution_failed, so the largest failure bucket becomes actionable.

How

classifyRunFailure already receives the run's events, so it reads the latest runtime_close diagnostic and promotes the generic execution_failed detail to the matching close reason. Only the opaque catch-all is refined — specific details (exit_code, cli_not_installed, …) and all retry behavior are untouched, so the change is observational-only. Three values added to TrackingRunFailureDetail.

This is the first step of the #3408 follow-up; mining the stream_error texts (1,082/wk carry a langfuse_trace_id) for real provider/tool error patterns is a Langfuse-backed follow-up PR.

Surface area

• API / contract — added stream_error / exit_nonzero / fatal_rpc_error to TrackingRunFailureDetail in packages/contracts
• UI / Keyboard shortcut / CLI / Extension point / i18n / new dependency
• Default behavior change — none for users; changes the failure_detail value emitted for a subset of failed runs (analytics only)

Validation

• pnpm --dir apps/daemon exec vitest run -c vitest.config.ts tests/run-failure-classification.test.ts — 44/44 (6 new red specs: each close reason → its detail, plus guards that a missing/unknown diagnostic stays execution_failed and a specific detail is never relabeled)
• pnpm --filter @open-design/contracts typecheck, pnpm --filter @open-design/daemon typecheck
• pnpm guard

Relates to #3408.

[PerishCode]

@lefarcen I reviewed the changed classifier path, the added daemon regression coverage, and the analytics contract enum extension. The implementation keeps the refinement scoped to the generic AGENT_EXECUTION_FAILED / execution_failed bucket, preserves specific process-exit details, and aligns the new failure_detail values with the existing runtime_close close-reason vocabulary. CI is green; my local focused test/typecheck attempt could not run because this prepared worktree has no installed dependencies. Nice work tightening the analytics signal without broadening runtime behavior.

_{🔁 Powered by Looper · runner=reviewer · agent=codex · An autonomous AI dev team for your GitHub repos.}