Session post-mortem subsystem + autonomous decision capture (A003)

[pskeshu]

Why

The first autonomous run (session 6a4a3d9b) was analyzed by hand and surfaced a stack of harness issues — 3 live embryos falsely terminated, a dead vision tool, and an empty decisions.jsonl that forced the whole analysis to be reverse-engineered from conversation.json. This PR turns that one-off into a repeatable capability and fixes the highest-leverage gap it exposed.

What

1. feat(postmortem) — Claude Code session post-mortem scaffold

Packages the forensic analysis into a repeatable, Claude-Code-authored, stored capability. Gently owns only the inputs/schema/storage/trigger; the intelligence lives in Claude Code.

• .claude/skills/session-postmortem/ — the analyst skill + report schema (v1)
• .claude/workflows/session-postmortem.js — fan-out → adversarial-verify → synthesize, parameterized by session
• docs/POSTMORTEM_DESIGN.md — architecture, storage layout, the dev/agent-facing feedback loop (recurrence-tracked rollup + regression evals), and the phased build plan

No harness runtime code changes in this commit.

2. fix(observability) — capture autonomous wake turns in decisions.jsonl (A003)

Autonomous (event-driven) turns wrote no Decision row — only the non-streaming call_claude path recorded one, hardcoded to trigger=USER_MESSAGE. So a fully autonomous run left decisions.jsonl empty.

• run_wake_turn now aggregates the turn's tool calls and writes one Decision(trigger=EVENT, trigger_detail=<wake reason>) per wake.
• _write_production_decision gains optional trigger/trigger_detail params; defaults preserve the USER_MESSAGE behaviour, so chat turns are unchanged.
• Capture is best-effort and never breaks a turn.

Scope: autonomous path only; user turns on the web streaming path remain a separate (smaller) capture gap.

Verification

• ✅ tests/test_wake_decision_capture.py (new, 4 cases) — EVENT trigger + wake reason + tool calls captured; default path still tags USER_MESSAGE; safe no-op when no decision log is attached.
• ✅ tests/test_eval.py (existing, 27) still pass — no regression.
• ⚠️ A full live wake turn through a running agent was not exercised here (needs the rig + API). Smoke-check on the instrument before relying on it.

Notes

• Based off feature/temperature-operations-all (not main), so the PR targets that branch.
• The first stored report (sessions/…_6a4a3d9b/postmortem/report.{yaml,md}) lives under D:\Gently3\ (runtime data) and is intentionally not in the repo.
• The report's item A005 (tool-arg coercion) is already fixed on the base branch by d846dcda.

Follow-ups (deferred; tracked in the report + design doc)

• Evidence-pack builder + put_postmortem storage
• _finalize_perception_run trigger + orphaned-run reconciler (A006)
• A004 (restore analyze_volume), A001/A002 (drift-aware termination + refocus write-back)

🤖 Generated with Claude Code