refactor(skill): split SKILL.md sections >500 lines into recipe files

[tirth8205]

Summary

Closes #313.

Anthropic recommends SKILL.md files stay under 500 lines (no guarantee that more is loaded into context). The /understand SKILL.md was 853 lines — well over target. This PR extracts large embedded bash/JS code blocks and JSON schema reference material into adjacent recipes/ and reference/ files referenced from SKILL.md, while preserving the full step-by-step workflow.

Files created

understand-anything-plugin/skills/understand/recipes/

• worktree-redirect.md (25 lines) — Phase 0 git worktree detection bash
• plugin-root-resolution.md (51 lines) — Phase 0 plugin discovery + first-time build bash
• understandignore-generator.md (30 lines) — Phase 0.5 .understandignore starter generator
• incremental-update.md (27 lines) — Phase 2 incremental update branch
• inline-validator.md (80 lines) — Phase 6 default-path validator script
• llm-review-path.md (26 lines) — Phase 6 --review dispatch block
• fingerprints-baseline.md (26 lines) — Phase 7 fingerprint baseline procedure
• intermediate-cleanup.md (17 lines) — Phase 7 cleanup bash

understand-anything-plugin/skills/understand/reference/

• knowledge-graph-schema.md (45 lines) — node/edge type tables, weight conventions
• output-shapes.md (61 lines) — layers[*], tour[*], assembled-graph JSON shapes

Each new file starts with a one-line header explaining its purpose and a cross-reference back to the relevant SKILL.md section.

SKILL.md line counts

File	Before	After
skills/understand/SKILL.md	853	533
skills/understand-knowledge/SKILL.md	132	132 (unchanged)
skills/understand-diff/SKILL.md	72	72 (unchanged)
skills/understand-dashboard/SKILL.md	105	105 (unchanged)
skills/understand-explain/SKILL.md	58	58 (unchanged)
skills/understand-onboard/SKILL.md	55	55 (unchanged)
skills/understand-chat/SKILL.md	55	55 (unchanged)
skills/understand-domain/SKILL.md	140	140 (unchanged)

The remaining 33 lines over 500 in the main SKILL.md are workflow narrative — the phased pipeline that needs to be read in sequence. Per the issue guidance ("a SKILL.md slightly over 500 doesn't need aggressive splitting if the structure doesn't naturally break apart"), I judged further fragmentation would hurt readability more than it would help context loading. All bash scripts, JS code blocks, JSON shape examples, and reference tables — the heaviest content — have been relocated.

Workflow semantics

All workflow semantics are preserved. Every see X reference points to a file in the same skill directory containing the exact code/content that was previously inline. The skill remains runnable end-to-end.

Test compatibility

Tests grep'd for hardcoded paths into the skill directory only reference the .mjs/.py scripts (compute-batches.mjs, extract-import-map.mjs, scan-project.mjs, merge-batch-graphs.py, extract-structure.mjs) — none reference SKILL.md content directly. The only test touching SKILL.md content is worktree-redirect.test.mjs, which inlines the bash snippet as a constant; I updated the doc comment in that test to point at the new recipe location, but the snippet content (and therefore test behavior) is unchanged.

Test plan

• pnpm install
• pnpm test (all suites pass — sandbox prevented me from running locally; please verify in CI)
• pnpm lint
• Manually verify the /understand skill instructions still flow logically end-to-end with the recipe cross-references

Closes #313

Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com