Add the five remaining build123d modeling skills

[SmartAI]

What this does

Second half of the skill-layer plan (#18 shipped the mechanism + sweep-and-loft): the five remaining build123d modeling skills, pure content on the existing machinery.

Skill	Teaches	Snippets
placement-and-frames	Plane/Location/Align datums, face-derived workplanes, numeric placement verification	align datum stack, side-face feature, placement probe
booleans-and-features	overlap-to-fuse, cutter overshoot, operation ordering, volume-delta verification	buried connector, through cutter, boolean probe
selectors-and-patterns	geometric selection (filter_by/sort_by/group_by), stale-selection rule, pattern-then-verify	fillet top edges, polar bolt circle, selector probe
holes-and-threads	Hole-drills-both-ways gotcha, census kinds (through/blind/internal), clearance vs nominal sizes, when real threads are worth it	counterbore, hole-directions demo, helix thread resource
surgical-edits	minimal-diff edits of gate-passed scripts, collateral-damage alarms, predict-then-verify deltas (worst CADGenBench pillar)	add-slot edit, volume-delta probe

Every skill follows the fixed template (when / invariants / recipes / failure signatures / go deeper), stays inside the 350-600 prose-word target (vitest caps at 800), and points down the disclosure chain at lookup_docs topics and search_docs queries instead of duplicating them.

SKILL_NUDGE_RULES gains four mappings (holes, selectors, booleans, placement; first match wins, sweep stays first). surgical-edits is deliberately unmapped - a regression is not detectable from one failure text. Every rule is vitest-checked against the registry.

All 15 snippets (10 new) execute against real build123d 0.11.1 in py-tests/test_skill_snippets.py, with volume/solid-count/topology assertions - including the hole-directions snippet that measurably demonstrates the both-directions gotcha, and the add-slot edit whose first draft accidentally proved the collateral-cut failure mode it warns about (the wall sat mid-plate; caught by its own assertion).

Validation

• npm run typecheck, npm test (399), npm run build green.
• py-tests/ 162 passed, all 21 snippet executions included.
• Full fake-LLM e2e suite green locally (CI does not run e2e).
• Prompt-budget test confirms the six-entry catalog stays under the 20k-char ceiling; full skill bodies still never enter the system prompt.

Human validation

The 4-condition ablation matrix (packages/client/eval/run-build123d-ablation.mjs) is now fully unblocked and remains the follow-up measurement.

🤖 Generated with Claude Code