perf(kernel): build each boolean once; fuse pattern instances in one boolean

[w1ne]

Problem

Evaluating a real part (a hinge with four 36-tooth patternCircular serration rings) took 6-10 minutes hosted and about 2-3 minutes on a Mac. Hosted render_preview then went over the MCP client's 300 s limit.

Root cause (measured)

A CPU profile of one 36-tooth ring showed 92% of the time in fusePatternInstances. That time split evenly between two calls: constructor$BRepAlgoAPI_Fuse_3 (30.0 s) and BRepAlgoAPI_BooleanOperation$Build (27.1 s).

1. Every history-aware boolean ran twice. The shape-taking constructors (BRepAlgoAPI_Fuse_3/_Cut_3/_Common_3) already run the operation. The Build() call after SetToFillHistory(true) then ran all of it again. This affects every union, subtract and intersect, not only patterns.
2. Patterns were folded pairwise. N instances took N-1 fuses, each one against the growing accumulator.

Fix

• runBooleanWithHistory default-constructs the builder, calls SetArguments, SetTools and SetToFillHistory, then builds once.
• New fuseManyWithHistory(body, tools[]) runs one general-fuse build over all inputs. The pattern lowerer uses it. Per-instance lineage retags are kept. mergeBooleanHistoryMany applies the input maps in order, so the first input wins a shared face, the same as the pairwise fold.
• booleanBuildStats.builds is a deterministic cost probe for tests.

Numbers

The reproduction part, run through the CLI built from this branch and from develop. Other jobs were loading the Mac (load average 18-30), so wall times vary.

command	develop	this PR
evaluate	118-203 s	47-103 s
export stl --parts all	145 s	39 s
export step	177 s	58-60 s
parts	109 s	43 s

The published kernelcad@0.15.0 on the same Mac took 146 s for evaluate, 104 s for STL and 83 s for STEP.

Geometry is unchanged. parts --json gives the same volume (7834.9954 / 14632.1058 mm³), surface area, bbox and triangle count (13312 / 6698) for both parts. Both STLs are watertight. The STEP file is the same size (2149749 bytes).

Tests

tests/unit/backends/occt/patternBooleanCost.test.ts counts OCCT builds, not wall time. On develop it fails: a 12-instance circular pattern costs 22 builds, a 5-instance linear pattern 8, and one pairwise fuse 2. On this branch each costs 1. A serration-ring case (teeth that touch along edges) checks the volume (= N × tooth), the bbox and the solid count. It passes on both develop and this branch, which shows the geometry is the same.

Local runs: tests/unit/{backends,naming,patterns,kernel,modeling,compute,capture,intent} plus src/kernel and src/modeling, 2692 passed. The 7 failures are in src/modeling/parts/* and fail the same way on develop: assets/parts was not generated locally.