feat: session observability & loop control (Phase 16)

[Catafal]

Implements PRD #3 in full — all four phases.

What's new

• Session radar: bach sessions list/adopt/resume/watch — discovers ALL Claude Code and Codex sessions on the machine (Bach-launched or not) with liveness states (running/idle/ended), task linkage, previews, filters and --json; REPL /sessions view.
• Codex parity: Codex launches now get sidecars; explicit bach hooks install <runtime> installers (parse-merge-rewrite, abort on unparseable settings, never clobber unrelated hooks). Supersedes the ADR-007 Codex exception (ADR-015).
• Observer: per-session run-and-exit watcher tailing transcript + hook-event spool; judge LLM (reuses run_codex_json + normalize_model) returning {summary, suggested_status, needs_human, confidence}; conservative authority enforced at the status write path — observer may only do executing→qa and {executing,qa,afk_queue}→hitl_queue, never done, never backward.
• Loop control (ADR-016): opt-in stop-hook verification gating (exit-2 continuation with failure output, max-continuation cap → hitl_queue), and a bounded headless AFK runner (max turns + wall-clock budget, never sets done).

Architecture

Two deep modules firewall all undocumented upstream knowledge: runtimes/session_discovery.py (on-disk stores + liveness) and runtimes/transcript.py (record parsing, never raises — unknown records preserved). Documented hook payloads are the control plane; transcript parsing only feeds previews/judge context, never gating.

Quality

• Built TDD by a 12-agent DAG workflow on strictly disjoint file sets
• 3-lens taste review: 3 STOP + 3 DRIFT + 3 WATCH findings — all fixed (watcher transcript-path placeholder, AFK overwriting hitl escalation, sessions watch wiring crash, frontmatter split corruption, installer overwrite-on-parse-failure, liveness age cap, shlex quoting, flock on spool, observability gaps)
• QA smoke against real machine session stores (radar found genuine orphan sessions; --state running correctly identified the live implementing session)
• Tests: 537 → 883 passed, gate green

Closes #3

🤖 Generated with Claude Code