[jonit-dev]

Follow-up to #644 (PRD-865), which the merge-queue landing shipped with a defect a third review pass then found. One commit, off develop at a6f32923b.

The defect

A spelled one-word name started at NAME_BASE_LEAD = 2, above the situation scores a prose request produces, so ordinary English that happens to be a symbol took row 1 — while the base commit (ecd2a1a56) answered all of these correctly:

request	merged (a6f32923b)	base	this PR
camera follows the player with a third person view	❌ View (a React rectangle) 2.26	✅ defineGame	✅ defineGame, View at 0.62
the enemy velocity should push the player back when hit	❌ velocity (three/tsl)	✅ gameplay rows	✅ velocity gone
make the water depth fog darker underwater	❌ depth (three/tsl)	✅ Atmosphere	✅ Atmosphere
make the scene feel foggy and moody	❌ Scene 2.95	✅ no wrong lead	✅ Scene gone
connect two rigid bodies with a hinge so the door swings	connect (WebTransport) in the top 3	Joint3D at row 2	connect gone

The fix

A one-word name spelled inside a longer request now scores (NAME_BASE_LEAD + NAME_SHARE_LEAD) * share, and must hold SINGLE_WORD_SHARE_FLOOR of the request and clear the name index's SYMBOL_WEIGHT_FLOOR. A request that is the name, and multi-word names, are untouched, so the names that should lead still do: add godrays, enable ssgi, use rapier, GPUParticles3D sparks, Area3D trigger, characterbody3d.

The four false leads above are now red-green tests in packages/engine-mcp/__tests__/search.spec.ts (they fail on a6f32923b, pass here), plus one keeping the names that must still lead.

Effect on the frozen held-out set

held out (50 rows)	base	merged	this PR
recall@1	0.400	0.620	0.660
recall@5	0.580	0.760	0.760
MRR	0.475	0.689	0.701
request-scope recall@1	0.200	0.400	0.400
mechanic recall@1	0.500	0.632	0.632

All four A/B gates pass; pnpm caps:recall stays 0.938053 (exit 0).

Also in this commit

• agent-files/.claude/hooks/discovery-gate.mjs and agent-files/.agents/skills/ponytail/SKILL.md still told an agent to open engine_capability_detail on every hit — the two places the router rule change missed. Both now read every row, and the sixteen scaffold trees are re-measured.
• The alias comment in capability-manifest.spec.ts named the wrong cause: SurfaceCache/RadianceCache are recalled through the name channel and GIBudget/ShadowFilterBudget through situations after the camelCase split — not through summaries.

Gates

pnpm exec vitest run packages/engine-mcp 158 passed / 1 skipped; packages/create-threenative/__tests__/scaffold.spec.ts green with the re-measured trees; instruction-budget, kit-conventions, capability-manifest, mcp-tool-schemas green; pnpm typecheck, pnpm lint, pnpm check:docs, pnpm sync:agents --check, pnpm quality green; pnpm caps:recall exit 0; scripts/capability-search-ab.ts --base ecd2a1a56 exit 0.