feat(mu): Agent MU Core — detect, reflect, retry

[gHashTag]

Problem

Pipeline fails silently. No error detection → reflection → retry loop. Competitors (Kiro, self-correcting agents) all have this. Our Link 11 (swe_fix) is a stub.

Solution

Agent MU — self-correcting pipeline agent:

1. Detect: After Link 7 (code_generate), run `zig ast-check` on output
2. Reflect: Parse error message, classify (type mapping, syntax, import, etc.)
3. Retry: Apply fix rule from MU Learning DB, re-generate, re-check
4. Max 3 retry iterations per spec

Architecture

```
Link 7 (gen) → zig ast-check → PASS → continue
→ FAIL → MU Detect
→ MU Reflect (classify error)
→ MU Retry (apply fix, re-gen)
→ Loop max 3x
```

Acceptance Criteria

• MU detects compilation failures after generation
• MU classifies error into known categories
• MU applies at least 3 auto-fix rules (missing import, type alias, reserved word)
• MU retries up to 3 times
• Failures logged to structured format

[gHashTag]

🧠 Agent MU Core — ⏳ QUEUED. Depends on pipeline PR merge. Next after #70.

[gHashTag]

🧠 Agent: Ralph | ⏱️ 2026-03-10T17:10Z
📋 Step: 0/5 — STARTING Agent MU Core
🔄 Status: THINKING

Thought: MU = Memory Unit. The swarm doesn't learn.
49 pipeline errors = 49 lost lessons.
MU will: detect failure → log pattern → suggest fix → retry.

Plan: 5 sub-tasks, sequential.

• MU-1: Research error pattern schema
• MU-2: Implement mu_agent.zig core
• MU-3: Implement .trinity/mu/ storage
• MU-4: Integrate Pipeline Link 11 hook
• MU-5: Test against 49 known failures

Next: Create sub-issues, start MU-1.