feat: skill-side deploy-gate CL-awareness (lockstep with tool side)

[jramos]

Summary

Ports the closed-loop-aware deploy gate to evolve_skill.py. Tool-side already shipped (CL-aware gate is on main); this PR brings skill side to schema v5 lockstep with the same trigger condition (band == weak_signal AND CL configured), the same _check_cl_primary_gate helper, the same v5 additive contract, and the same abort-path discipline.

Behavior

• weak_signal band: runs closed_loop_cache.force_run(evolved_body) post-GEPA, gates on _check_cl_primary_gate (CL gain ≥ growth-scaled required + synthetic regression within tolerance) plus the preserved _check_absolute_char_ceiling (wallpaper protection).
• All other bands (healthy, no_headroom, uniform_failure): today's synthetic-only gate path runs unchanged.
• --no-saturation-check: falls through to synthetic gate; gate_decision.json records reason_synthetic: "preflight_skipped".

Skill-specific differences from the tool-side equivalent

• force_run(evolved_body) — NOT force_run(evolved_full). The cache was constructed with baseline_artifact_text=skill["body"], so the deploy-time call must use the body (no YAML frontmatter) for the cache key to match the preflight baseline. Otherwise we silently double-spend on the evolved eval. Test 11 (test_force_run_called_with_skill_body_not_full) pins this contract with belt-and-suspenders substring negative checks (--- / name: absent from the call arg).
• Abort paths write evolved_FAILED.md (not .json) — matches the existing skill-side rejection convention. Test 12 pins this.
• V4 skill-specific payload fields (bap_max_growth, bap_safety_margin, eval_source, fitness_profile, proposer_mode, knee_point.band_roster[*].holdout_score) all preserved in v5. Test 13 pins this.
• growth_pct computed on evolved_full vs skill["raw"] (matches existing skill-side validate_growth_with_quality convention).

Schema v5

Additive over v4. All 4 gate_decision.json write sites in evolve_skill.py now emit v5: static-fail, cl_eval_failed abort, cl_eval_incomplete abort, main success/reject. write_cost_ceiling_abort() call also bumped via the schema_version="5" kwarg added in PR #69. New fields surface only when CL-primary fired (same shape as tool side).

Commit sequence

1. refactor(evolve_skill): preserve SaturationReport fields for deploy gate — plumbing, no behavior change
2. feat(evolve_skill): branch deploy gate on saturation band — main behavior change + run_inputs hoist
3. feat(evolve_skill): gate_decision.json schema v5 with CL-primary fields — payload extension
4. test(evolve_skill): integration + schema regression for CL-aware gate — 13 tests
5. test(evolve_skill): pin CL-primary path in absolute-char-ceiling test — review feedback fix (the absolute_char_ceiling check turned out to actually live in the CL-primary branch, not in static gate — corrected the test to pin the verified production flow)
6. test: weakened systematic-debugging fixture for CL-aware-gate smoke — empirically-probed fixture that lands in weak_signal band

Empirical fixture-design finding

Task 5 surfaced a non-obvious empirical insight worth documenting: passive weakening (stripping methodology sections) of a skill does NOT move validator behavior on systematic_debugging.jsonl — gpt-5-mini gets 5/5 regardless of skill content because the user_message is explicit. gpt-5-nano gets 0/5 with any weakening. The only path to weak_signal was active misdirection — a skill that LOOKS plausible to the LLM judge (synthetic ≥ 0.95) but redirects the agent toward different behavior (write a report instead of editing). This is the same binary-tier issue that PR #68 worked around for the tool side; the active-misdirection technique is the skill-side analog.

Test plan

• uv run pytest -q — 1130 passed locally (+13 from this PR: 10 tool-side analogs + 3 skill-specific guards)
• CI green across 4 Python versions
• Optional manual smoke (~$3–5): uv run python -m evolution.skills.evolve_skill --skill weakened-systematic-debugging --skill-source-dir tests/fixtures/skills --closed-loop-during-evolution evolution/validation/suites/systematic_debugging.jsonl --closed-loop-hermes-repo …/hermes-agent --closed-loop-agent-model openai/gpt-5-mini --iterations 10 --seed 42 --max-total-cost-usd 15 --closed-loop-mode trainset --closed-loop-in-valset --force-saturation-check --gepa-minibatch-size 8. Expected: decision: "deploy", decision_signal: "closed_loop".

Scope notes

• Tool-side and skill-side now at lockstep schema v5.
• Follow-up flagged: run_inputs literal is duplicated across evolve_skill.py and evolve_tool.py. After this PR there are 6 sites across 2 files. Worth extracting to evolution/core/run_inputs.py in a separate cleanup PR; not in scope here.
• Memory-worthy finding (post-merge): the active-misdirection weakening technique for test_command-based skill suites. Add a project memory entry so future skill-side empirical work doesn't waste time on passive weakening that won't move validator behavior.