A full brick build disagrees with its own raw field inside the band -- up to 0.263 on random documents, and a Move leaves those bricks stale #649

A full brick build of a document disagrees with that document's own raw field (clay_eval_points) at in-band samples. This is on main, before #648; #648 only made it reachable from undo.

Measured

benchmarks/undo_bound_oracle_probe with RV_RAWCHECK builds each random document's brick cache from nothing and compares every in-band (|raw| <= 0.12) sample of a built brick with clay_eval_points at the same point, flagging anything off by more than 2e-3 (far above the fp16 step at that range). Probe linked against main at c62d521 (#646 merged, before #647/#645/#648), and again against main + #645/#647/#648 -- identical numbers:

The seeds #648's review named:

Why it matters: the forward Move already leaves these bricks

RV_FORWARD does a host Move segment, dirties exactly what clay_layer_move_surface_regions reports, refills, and compares with a rebuild on a saved copy. On main:

• seed 933 (plain): 8 bricks differ
• seed 5229 (plain): 3 bricks differ
• seed 5229 (rich): 6 bricks differ

Before #648 an undo of that Move refilled the whole node, which happened to hide the disagreement. #648 narrows the undo to the grab's support, so the undo now leaves the same bricks the Move does: 5 of 3,000 random undo/redo trials (1,500 plain, 1,500 rich), 1 to 8 bricks each (a sixth, 5128, leaves 13 on main too).

Which side is wrong

#648's review checked each of those bricks: the values the narrowed refill KEPT match the raw field to the fp16 step; the full REBUILD is the one that is off. So the defect is in the build path (cull / seed / brick fill of a whole document), not in the undo bound or in the Move's reported regions.

Reproduce

cmake -S . -B build/probe -DCLAY_BUILD_BENCHMARKS=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build/probe --target undo_bound_oracle_probe --parallel 8
RV_RAWCHECK=1 build/probe/undo_bound_oracle_probe 933 1        # plain
RV_RAWCHECK=1 build/probe/undo_bound_oracle_probe 5111 1 r     # rich
RV_FORWARD=1  build/probe/undo_bound_oracle_probe 933 1
RV_FORWARD=1  build/probe/undo_bound_oracle_probe 5229 1 r
RV_RAWCHECK=1 build/probe/undo_bound_oracle_probe 1 1500       # the sweep

The probe drives the C ABI only, so it links against any revision for an A/B. RV_VERBOSE traces the fixture. Background: openspec/changes/archive/2026-09-23-bound-an-undone-grab-by-its-support/tasks.md, "What review found".