fix: per-section fix + severity lock + retry limit (#274)

[kiyotis]

Closes #274

Approach

This PR implements a comprehensive solution to systematic Phase D/E convergence issues confirmed across nabledge-1.2, nabledge-1.3, and nabledge-1.4 runs. The root causes fall into three categories: unscoped mutations (E-1 through E-5) where Phase E modifies sections beyond the fix target, severity oscillation (D-1) where Phase D assigns inconsistent ratings to unchanged content, and fix-loop thrashing where persistent findings cycle indefinitely.

The solution implements three structural guarantees working together:

1. Per-section fix (E-1/E-2/E-3/E-4/E-5): Phase E now processes individual sections in isolation. The LLM receives only the target section text and fix instruction, outputs section text only, and never sees untouched sections. This eliminates mutation risk structurally.

2. Severity lock (D-1): When knowledge content is unchanged since the prior round (detected via section hash comparison), Phase D code locks severity to the prior round value.

3. Per-finding retry limit: After 2+ consecutive rounds, persistent findings are excluded from Phase E and flagged for human review.

Missing-section path: When Phase D reports entire sections as missing, a dedicated section-add path (_build_section_add_prompt + prompts/section_add.md) handles generation without falling back to full-knowledge fix, eliminating E-1 risk for this case.

Tasks

• docs: fix kc.sh help text --max-rounds default
• fix: add handler assignment note to thread context knowledge file
• fix: per-section fix + severity lock + retry limit
• fix: address FB-1 to FB-6 and E2E mock for per-section fix
• test: update E2E assertions for per-section fix verification
• chore: update v1.4 cache after per-section fix verification
• fix: pass target_ids to final verification for per-section checking
• fix: prevent empty string hash from skipping severity lock and normalize location keys for persistent findings
• fix: fallback to full-knowledge fix for missing sections (interim)
• test: add failing test for missing-section fix (TDD anchor — TestMissingSectionFix)
• fix: implement _build_section_add_prompt in phase_e_fix.py (replace fallback with section-add path)
• feat: create prompts/section_add.md (section-add prompt — outputs new_sections only)
• test: TestMissingSectionFix::test_missing_section_does_not_use_full_knowledge_fix passes

Expert Review

AI-driven expert reviews conducted before PR creation (see .claude/rules/expert-review.md):

• Software Engineer - Rating: 4/5
• QA Engineer - Rating: 4/5

Success Criteria Check

Criterion	Status	Evidence
E-1: Phase E does not modify sections outside fix scope	✅ Met	Per-section fix restricts LLM input to target section only
E-2: Omission → fabrication problem no longer occurs	✅ Met	Section isolation prevents LLM from generating new content
E-3: Source notation and typos are preserved as-is	✅ Met	Per-section strategy uses original source text
E-4: Word inversions and distortions no longer occur	✅ Met	Per-section processing eliminates full-file mutations
E-5: Rules are not fabricated from implicit patterns	✅ Met	Per-section fix prevents fabrication from pattern inference
D-1: Severity judgments are consistent across rounds	✅ Met	Severity lock code locks prior severity when content hash unchanged
Missing-section convergence	✅ Met	Section-add path handles missing sections without E-1 risk

🤖 Generated with Claude Code