Shared movement/cost geometry consolidation

- Removed estimate/geometry.rs and low/expand/cache.rs. No selected unit-buffer copy recipes remain.
- Copy family owns descriptor selection/coalescing from shared matched rows and actual alias identity.
- Destination coverage does not inspect unnecessary source traversals for physical/panel copies.
- Workspace tests: 436 passed, 8 ignored. Includes full cached/disabled expanded-program comparisons, symbolic coverage, mixed-layout byte mappings, and actual backing-alias ordering.
- Representative MLP and small FP8 ViT packages compared against bound-copy-calls artifacts.
- Larger benchmark structural counts, costs and row estimates match: structural-equivalence.json. The 27-layer comparison uses a fresh baseline on both compilers, because obsolete checkpoint migration was removed.
- Capacity-based retained geometry estimates: MLP 70.96 MB -> 47.51 MB; attention 25.11 MB -> 3.91 MB; fresh 27-layer 183.81 MB -> 33.61 MB. These estimates include shared traversal overcounting.
- Timing samples are mixed, not a demonstrated speedup: MLP warm expansion 0.95 s -> 1.54 s and footprint 0.35 s -> 0.69 s; attention warm expansion 0.42 s -> 0.43 s; fresh full model cold expansion 5.95 s -> 6.15 s, warm 6.40 s -> 10.05 s. Full-model recosting and unchanged frontend planning also varied substantially. Raw before/after JSON reports are retained.