feat(skills): periodic heuristic promotion from ERL to full skills (AutoSkill A6)

[bug-ops]

Description

ERL (erl.rs) already accumulates per-skill heuristics over sessions via Jaccard deduplication. Add a periodic aggregation pass that examines skills where the heuristic count exceeds a configurable threshold (e.g. 5 unique heuristics) and asks the LLM whether any combination of heuristics is substantial enough to:

• Become a standalone new skill (novel capability distinct from parent)
• Be merged into the parent skill body (enriching the existing skill)

This closes the loop between short-term ERL learning and the long-term skill corpus.

Motivation

AutoSkill's lifelong learning model distinguishes itself by making learned behaviors permanent. Zeph's ERL heuristics are valuable but ephemeral — they enhance prompts per-session but never graduate into durable skills. Periodic promotion creates a compounding improvement loop.

Acceptance Criteria

• skills.heuristic_promotion config section: enabled, threshold (default 5), provider, interval_hours
• Periodic job (scheduler or background task) scans skills with heuristic count >= threshold
• LLM eval prompt: given N heuristics for skill X, classify as Standalone/Merge/Keep
• Standalone: creates draft skill (user review required, same as A1 flow)
• Merge: triggers versioned merge flow (A2) with parent skill
• Keep: resets counter, tries again after another N heuristics accumulate
• Disabled by default

References

• AutoSkill paper: https://arxiv.org/abs/2603.01145
• Research evaluation: specs/061-autoskill-heuristic-promotion/spec.md
• Related: feat(skills): conversation trace → skill extraction pipeline (AutoSkill A1) #4447 (draft skill flow), feat(skills): versioned skill merging with Add/Merge/Discard decision flow (AutoSkill A2) #4448 (versioned merge flow)