bf-4b1c wants test coverage proving geometry placement drives fusion peaks,
with the explicit "would fail with old co-located placement but passes with
new geometry" property. The prior children (bf-18yn -> bf-3y9r) cover each
half separately — TestEngine_DefaultPlacementProducesPeaks (spread alone,
geometry-dependent synthetic links) and TestEngine_CoLocatedOriginYields
NoPeaks (co-located alone). What was missing was a single run that isolates
placement as the SOLE variable and demonstrates the non-zero-peak condition
flipping with geometry, proving the test is genuinely regression-sensitive.
Add TestEngine_GeometryPlacementDrivesFusionPeaks: a fuseWithPlacement helper
builds an engine sized to DefaultSpace, seeds N nodes at a given placement,
and fires an explicit geometry-independent motion link for EVERY node pair
(DeltaRMS=1.0, Motion=true) — the link set, grid, and room are identical
across both legs, so only where nodes sit differs. For default (spread)
placement the grid accumulates a non-zero max (gridMax>0); for the old
co-located (0,0,1) collapse every link is degenerate (<0.1 m), AddLink
Influence early-returns, and the grid stays at zero (gridMax==0). The
differential asserts spread>0 && colocated==0. For count>=4 the crossing
links additionally extract >=1 blob, confirming the non-zero grid localizes.
The trace under `go test -v` shows the flip:
[spread] nodes=4 gridMax=1.0000 blobs=2 top=(2.90,2.50,1.30) conf=0.768 activeLinks=6
[colocated] nodes=4 gridMax=0.0000 blobs=0 activeLinks=6 — degenerate links, no accumulation
go test ./... , go vet ./... , and go build ./... (all three modules) pass.
Co-Authored-By: Claude <noreply@anthropic.com>
c4d42e8c30