feat(loop): add .learnings/ continuous error capture loop with threshold promotion

[harshsinghmp]

Classification

actionable-feature

Context & Rationale

To prevent recurring agent failures across multi-agent sessions, we need a standardized .learnings/ feedback loop that captures corrections, errors, and feature requests, automatically promoting durable lessons to workspace memory.

Key Capabilities & Scope

1. Support .learnings/ Structure:
   • LEARNINGS.md: Corrections, insights, knowledge gaps, best practices.
   • ERRORS.md: Tool/command failures, reproducible traces, sanitized contexts.
   • FEATURE_REQUESTS.md: Requested capabilities.
2. Deterministic Entry ID Scheme:
   • [LRN-YYYYMMDD-XXX], [ERR-YYYYMMDD-XXX], [FEAT-YYYYMMDD-XXX].
3. Automated Promotion Engine:
   • Threshold Rule: When an entry has recurrence_count >= 3, seen across ≥2 distinct tasks within a 30-day window, automatically promote it to CURRENT.md and workspace AGENTS.md as an actionable prevention invariant.

Verification Criteria

• CLI / MCP tools support capturing into .learnings/.
• Automated promotion fires when count hits threshold.
• Passes test suite (bun test test/).