[wass08]

What does this PR do?

With the usage checks on (setSdfChecks(true)), verify_scene took every usage measurement on every call, whatever type it asked for. On a large building the circulation measurement (passages, turning circles, pockets) is nearly all of the cost, so a question about skirting or a blocked door paid for it too, and a run_batch of several typed checks paid for it once per call.

Two changes in packages/core:

1. A typed question takes only the usage measurements its type needs. usageIssues now lists its measurements with the finding types each gives (types option), and the usage check passes the type asked for. A usage type takes its own measurement; any other type takes none; usage_room takes the one behind the findings a room entry holds. A misspelt usage type still takes all of them, so its answer names the types there are.
2. The path solver's heap allocates nothing per pop. Same comparisons in the same order, so the same paths; it only removes the per-step temporaries.

Measured on a saved six-storey residential scene (1,244 nodes) on an Apple-silicon laptop, usage checks on:

Call	Before	After
Full check (no type)	15.7 s	11.2 s
type that is not a usage finding (door_blocked, rooms_no_skirting, item_overlap, …)	14.3–15.3 s	0.3 s
usage_clash, usage_opening_blocked, usage_bed_clearance, usage_fixture_clearance, usage_door_over_stair	13.9–15.3 s	0.3–0.4 s
usage_door_swing	14.8 s	1.4 s
usage_entry_narrow	14.4 s	1.6 s
usage_passage_narrow, usage_pocket, usage_turning_narrow, usage_room	14.9–15.3 s	8.0–8.9 s

The circulation types still cost seconds: that measurement is where the time goes, and this PR only stops other questions from paying for it.

Results. On that scene the full report and the model's reading of it are byte-identical before and after, and so are the findings and their count for each of the 15 types asked. One thing differs: an answer with no findings lists the other types found (types), and that list now holds the types of what the question measured. For a type that is not a usage finding it no longer lists usage types.

How to test

1. cd packages/core && bun test src/agent-operations/verify-scene-typed.test.ts: on a small two-room scene whose report holds six usage types beside the ordinary ones, each typed answer equals the findings of that type in the whole report; a room with a type answers the same; a misspelt usage type answers with every type; the report for a non-usage type holds no usage findings.
2. bun test src/spatial/usage-field src/agent-operations src/agent-tools (1,806 pass) and, in packages/mcp, bun test src/tools/verify-scene.test.ts src/tools/scene-query.test.ts (12 pass).
3. To see the timings: load a large multi-storey scene, call setSdfChecks(true), and time AGENT_OPERATIONS.verify_scene(nodes, { type: 'door_blocked' }, context) against {}.

Screenshots / screen recording

N/A — non-visual change

Checklist

• I've tested this locally with bun dev
• My code follows the existing code style (run bun check to verify)
• I've updated relevant documentation (if applicable)
• This PR targets the main branch

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes which usage measurements run for typed queries and how empty-result metadata lists types; heap refactor must preserve path results but is covered by existing usage-field tests.

Overview
When verify_scene is called with a type, usage checks no longer run the full circulation and fixture suite on every request. usageIssues maps each measurement to the finding types it produces and runs only matching measurements; non-usage types skip usage work entirely, usage_room still folds to pocket/turning types, and unknown usage types fall back to measuring everything (so answers can list valid types).

The usage scene check wires this through askedTypes(input.type) before calling usageIssues.

The grid path Heap is also tweaked so pop returns only the cell index and exposes the priority via heap.key, avoiding per-step tuple allocation in widestPath, travel, and shortestRoute.

A new verify-scene-typed.test.ts asserts typed reports match the full report per type, room+type behavior, misspelling fallback, and that non-usage types emit no usage findings.

^{Reviewed by Cursor Bugbot for commit 2d5dc8a. Bugbot is set up for automated code reviews on this repo. Configure here.}