feat: phase-2 harness — feature map, fixture packs, verdict replay

[aisona-lab]

Summary

Phase 2 harness-only work for lazycoder:

• docs/FEATURE_MAP.md — maps core / eval / CLI surfaces to tests and fixture packs; linked from README.
• fixtures/{clean,deny,empty,malicious,boundary}/ — eval case bodies rehomed from config/evals.json. The config file is now a thin fixture_packs + scoring wrapper (loader refuses embedded cases to avoid dual-source drift). Fixtures are bundled in the wheel next to config_defaults.
• lazycoder replay LOG — recomputes derive_verdict from each decision-log JSONL record with no model call; exit 1 on drift. Documented in README; covered by pytest (also exercised via the existing decision-log replay unit test).
• README — short sibling section: lazycoder = trailer, agent-action-gate = thesis (no dependency claimed).

Out of scope (unchanged)

Stage 2 corpus, Action fail-on default, rubric/policy taste, LLM prompts/models, SARIF, sandbox.

Test plan

• uv sync --extra dev && pytest -q (111 passed)
• ruff check . && black --check . && mypy src
• Manual lazycoder replay on a fake-client decision log
• CI green on this PR

[github-actions]

lazycoder review — 💥 ERROR

Downloading anthropic (1.3MiB)
Downloading pydantic-core (2.0MiB)
 Downloaded pydantic-core
 Downloaded anthropic
Installed 16 packages in 37ms
error: Error code: 400 - {'type': 'error', 'error': {'type': 'invalid_request_error', 'message': 'Your credit balance is too low to access the Anthropic API. Please go to Plans & Billing to upgrade or purchase credits.'}, 'request_id': 'req_011CfXa7Y4HXKLgsLuQ5dnE6'}

_{0 finding(s), 0 abstention(s), 0 rule error(s) — R1..R17 rubric via lazycoder}