# Explicit value views: verification

Base commit: `15ce9043`.

Low logical values now retain concrete `ShardView` bindings. Copy reuse updates those bindings directly; geometry and kernel construction no longer repair borrowed shard IDs afterward. Logical shape remains separate from backing strides. Source ownership and diagnostics preserve the selected view. Writes, Sum reinterpretation and Repeat retain their checked whole-allocation requirements.

Validation on 2026-09-15:

- Release workspace tests: 440 passed, 0 failed, 8 ignored. Log: `/tmp/ipu-resolved-reads-workspace.log`.
- SDK exchange replay: 1 passed. Log: `/tmp/ipu-resolved-reads-replay.log`.
- `cargo check --workspace --all-targets`, `cargo fmt --all -- --check`, and `git diff --check` passed.
- Production source contains no `borrowed_views`, `resolve_read_view`, or `value_shards` path.
- F16 MLP Repeat3 package is byte-identical to the committed baseline (0.321 s build, 48,092 KiB peak RSS).
- Small FP8 ViT Repeat3 package is byte-identical (0.963 s, 209,224 KiB).
- Saved full 27-layer SigLIP package is byte-identical (107.882 s, 4,311,344 KiB).
- Commands, result JSON and logs for package comparisons are in this directory. These are build/comparison observations, not controlled compile benchmarks.

Regression coverage includes borrowed slices and broadcasts, source selections sharing one backing shard, factor mappings, canonical alias/Repeat/reduction inputs, unmaterialized input rejection, and diagnostic logical indexing with backing row strides. No identical hardware benchmark was rerun. The preceding typed-attention validation ran six real images with pretrained weights against FP32 reference (minimum cosine 0.994567066); its device outputs were byte-identical to the earlier validated package.

Remaining structural work: shared alias-relative access geometry (root plus signed origin, physical interpretation and selection), common local-copy binding, movement policy ownership, cache consolidation, and the final comprehensibility/production-size audit. This change does not finish that work.

Preliminary source-size measurement in `../source-size/` excludes test-only Rust items, test paths and comments. It reports 48,402 implementation lines at proposal baseline `1501698`, 50,818 at `15ce9043`, and 50,846 in this worktree: +2,444 (about 5.0%). Classification still needs auditing; net code-size reduction is not achieved.