feat: --gepa-minibatch-size CLI flag (Path E)

[jramos]

Summary

• Exposes GEPA's existing reflection_minibatch_size kwarg as --gepa-minibatch-size on both evolve_skill and evolve_tool. Default unchanged at 3 (matches GEPA's own default; no behavior change for existing scripts/CI). Users hitting the saturation pre-flight's weak_signal band now get the panel telling them to bump to 8 + pipeline-specific budget compensation.
• New EvolutionConfig.reflection_minibatch_size field (canonical home, auto-serialized into metrics.json).
• Post-dataset-build trainset-ceiling guard that aborts at startup with an actionable message if --gepa-minibatch-size exceeds the trainset size. Without the guard, GEPA's EpochShuffledBatchSampler asserts mid-optimization at gepa/strategies/batch_sampler.py:71.
• Updated saturation_check.py's weak_signal suggestions to recommend the new flag concretely (replaces the "Path E follow-up would help once landed" placeholder).
• Help text is pipeline-aware: tool side recommends --iterations bump (uses max_full_evals); skill side recommends --budget heavy (uses auto).

Background

Implements Path E from reports/pareto_frontier_feasibility.md. Spike #2 showed that on a saturated baseline with ~15% behavioral failure rate, GEPA's sum(subsample_scores) acceptance gate rejected all 40 candidate proposals over 116 iterations — because a random 3-example minibatch contains a discriminating example only ~34% of the time (hypergeometric, K=7 / N=56). Bumping to 8 raises that to ~68%. Path E ships the user-facing knob; Path F (saturation pre-flight, already merged) is the user-facing surface that recommends it.

Test plan

• Pull, run uv run pytest -q — expect 1082 passed (up from 1078 pre-branch, +4 new tests across TestGepaMinibatchSizeFlag in both pipelines).
• Inspect help text: uv run python -m evolution.tools.evolve_tool --help | grep -A 12 gepa-minibatch shows the tool-side help mentioning --iterations; uv run python -m evolution.skills.evolve_skill --help | grep -A 12 gepa-minibatch shows the skill-side mentioning --budget heavy.
• Invoke with --gepa-minibatch-size 1000 against a tiny synthetic dataset; expect a clean exit 1 with "exceeds trainset size" in stderr instead of a deep GEPA assertion.
• (Optional, ~$35) Multi-seed smoke against the spike fix: bump DSPy 3.0→3.2 and make GEPA actually run #2 saturated baseline: 3 seeds at --gepa-minibatch-size 8 (success: ≥2 of 3 produce at least one accepted proposal) + 1 reverse-control at --gepa-minibatch-size 3 (success: reproduces the spike fix: bump DSPy 3.0→3.2 and make GEPA actually run #2 all-rejected pattern). Verifies the mechanism actually moves selection.

Scope notes

• Path D (Pareto-dominance acceptance) and Path C (stratified sampling) remain future work. Path C is shippable without an upstream PR via dspy.GEPA(gepa_kwargs={"batch_sampler": StratifiedBatchSampler(...)}) if Path E proves insufficient on harder cases.
• Type-design refactor from PR feat: saturation pre-flight for evolve_skill and evolve_tool #64 review (bundle closed_loop_* triple, DEFAULT_THRESHOLDS → frozen dataclass) is a separate follow-up, not in this PR.