Add eval system Phase 3 — judge runner end to end ...

Fills in the stubs from Phase 2:
- judge.render_session: full transcript with tool_call/tool_result folding,
  reactions inlined per assistant block, planning_logs appendix, no
  compression-summary substitution
- judge.run_expert: real LLM call, fence-tolerant JSON parse, single retry
  with corrective nudge on schema or parse error
- judge.evaluate_session: asyncio.gather across the three experts
- db.EvalDB: insert_evaluation_run (txn), list_evaluations,
  evaluated_session_ids, feedback_by_index helper
- cli `run` (filters: --session, --since, --limit, --re-evaluate-all,
  --dry-run, --model, --backend) and `show` (groups by eval_run_id, prints
  per-expert axes plus averaged scores)

Verified end-to-end against a real 10-message secretary session:
all three experts returned valid JSON first try; spread between strict
critic and the others surfaced as expected.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Add eval system Phase 2 — rubric, expert prompts, judge skeleton ...

Drafts the v1 rubric (7 axes, anchors at 10/30/50/75/100, open scale),
three independent expert prompts (strict_critic / pragmatist / tech_lead)
that all return the same JSON shape, and the orchestration scaffolding:
schema.py (pydantic models), judge.py (rubric loader, score averaging,
fence-tolerant JSON parser, new_run_metadata), cli.py with argparse for
run / show / stats. Real LLM calls and transcript rendering land in
Phase 3 — the stubs raise NotImplementedError.

`python -m debug.eval` works as the entry point. Anchor `examples` are
left empty for now; user fills them with real session_ids later without
bumping rubric_version.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>