feat(forge): SkillOpt-style epoch training with mini-batch validation

[criptogus]

Summary

Brings ideas from microsoft/SkillOpt ("train agent skills like neural nets — epochs, mini-batches, validation gates, without touching weights") into our existing SkillForge loop. Our autoLearnPipeline already had evolutionary search (generations + elite archive + A/B fitness), so this PR adds the missing loop-level training mechanics rather than importing the Python project.

Phase 1 — epoch training + mini-batch validation (runForgeLoop)

• New inputs: epochs (1–6), batch_size, candidates_per_gen.
• Each epoch runs eval → auto-learn → validates the candidate patch on a rotating mini-batch of golden cases.
• Maintains a single best_skill snapshot (the running champion); improvements compound across epochs.
• Promotion is gated: the champion is only persisted as a new version if it does not regress on the full golden set (loop:after).
• Per-epoch history + training metadata returned and stored in evolution_trace.

Phase 2 (groundwork) — trajectory-driven signal

• buildExecutionSignal summarizes real agent runs from skill_executions (success rate, top error kinds, per-model success) and feeds them into autoLearnPipeline as the strongest failure evidence for root-cause analysis and patch proposals.
• Note: full step-by-step trajectory capture (input/output per step) would require a schema migration; this uses the outcome telemetry available today.

Notes

• Backwards compatible: defaults (epochs=1) reproduce the previous single-pass behaviour. patch in the result is now nullable (no scorable candidate) — UI guarded.

Test plan

• Run forge loop with epochs=1 on a package — verify before/after parity with prior behaviour.
• Run with epochs>1 — verify epoch_history, mini-batch validation, and best-snapshot promotion.
• Verify a regressing champion is rejected by the full-set gate.
• Confirm execution telemetry appears in evolution_trace when skill_executions exist.

https://claude.ai/code/session_01EntkmBiYh381pKqvvSFBkg

---

Generated by Claude Code