Movement geometry/policy/cache verification, 2026-09-14.

Destination geometry retains symbolic coverage, evaluated on demand, and has no
packing decisions. Each mid Copy carries a packing policy. Default construction
retains Automatic. Forcing a policy retains that site's copy boundary; unsupported
physical requests fail expansion. The selector declares actual scratch and work.

- Workspace release suite: 423 tests passed, 7 ignored.
- Codegen release suite: 318 tests passed, 5 ignored; one doctest passed.
- Byte interpretation tests execute direct and staged copies through padded packed
  layouts and compare the values. Staged work must declare extra storage; subword
  requests unsupported by direct exchange are rejected instead of silently staged.
- Coverage tests compare exact uncovered bytes against an independent bitmap.
- Cache tests cover collisions, growth with supplied hashes, bounds, ownership
  independence and precision-sensitive byte counts. Expansion tests compare entire
  cached/uncached low graphs, including Repeat.

Hardware/reference checks (small randomized fixtures, 64 selected compute tiles):
F16 MLP Repeat3 passed, maximum absolute error 0.000183;
FP8 ViT with 3 layers and fused QKV passed, maximum absolute error 0.189453.
Modeled final total/exchange cycles remain 32613/18087 and 331874/120708.
Hardware commands, logs and binary profiles are adjacent. These runs preceded
the lazy evaluation of the same pure coverage, consolidation of single-copy
and batched-copy dispatch, and additional tests. Final codegen tests passed after
those changes; no device instruction implementation changed.

Expansion benchmark, pinned to eight CPU cores, one sample per variant. This
excludes search time, placement and exchange scheduling. Host timings are noisy.
The before binary is commit 675d454; commands retain its exact executable path.

| Workload | Before cached (ms) | After cached (ms) | After uncached (ms) | Expansion RSS before/after/uncached (MiB) |
|---|---:|---:|---:|---|
| mlp-b2-repeat3 | 1474.4 | 829.6 | 967.6 | 212.1 / 205.7 / 162.0 |
| attention-materialized-b1 | 548.9 | 462.4 | 1411.4 | 128.4 / 121.0 / 95.2 |

All recorded work counts, low cycle estimates, row-footprint estimates and
geometry counts match between these three variants. Earlier before/after samples
are also retained: the first implementation eagerly scanned unnecessary coverage
and was slower. The final implementation leaves that coverage symbolic until a
clear needs it. These samples support retaining the distinct copy/geometry caches;
they do not establish an end-to-end compilation speedup.