Why

bgz-tensor is excluded from CI — no workflow tests its manifest (build.yml/rust-test.yml/style.yml only cover lance-graph, lance-graph-contract, deepnsm, jc). As a result cargo test -p bgz-tensor was 195 passed / 5 failed on main, and had been for an unknown number of PRs. Surfaced by a local sweep of the CI-uncovered standalone crates (the same gap that hid the causal-edge red fixed in #498).

Each red was diagnosed as test-stale vs code-bug before fixing — no assertions were blanket-bumped to green (which would have papered over the real defect).

The 5 failures

Test-stale (the code was correct; the test's expected value had drifted):

Code bug (a genuine robustness defect, not a stale test):

• matryoshka::{encode_decode_roundtrip_nonzero, roundtrip_quality_reasonable} panicked with an out-of-bounds slice when the SVD basis rank is lower than the band profile's nominal max — e.g. building SvdBasis::build(rows, 128) from only 50 sample rows yields a rank-50 basis, but BandProfile::standard(128, …) has bands starting at 64/192. encode_row sliced coeffs[band.start..] with an unguarded start; decode_row wrote coeffs[band.start + i] past the buffer.

  • Fix (symmetric, both sides): encode clamps the slice start (it was only clamping end); decode caps the per-band component count to the coeff buffer (n.min(coeffs.len().saturating_sub(band.start))). A band entirely beyond the available rank now degrades to empty on both sides, so the byte stream stays in sync — the roundtrip + quality assertions (cosine > 0.8, len == 256) pass.

  This one was a latent crash, not merely a stale test: any production use with a rank-deficient basis (fewer sample rows than requested components) would panic.

Verification (local — this sandbox has protoc + full toolchain)

cargo test  -p bgz-tensor                        → 200 passed / 0 failed   (was 195/5)
cargo fmt   -p bgz-tensor -- --check             → clean
cargo clippy -p bgz-tensor --all-targets -- -D warnings → clean

Diff is 9 insertions / 6 deletions across 4 files — the three one-line test corrections (each with a corrected explanatory comment) plus the two-site matryoshka guard.

Follow-up (not in this PR)

The root cause is structural: CI tests only 4 of ~30 crates. Two CI-invisible reds have now been found and fixed this session (causal-edge #498, bgz-tensor here). A separate PR should add the disk-cheap standalone crates (bgz-tensor, bgz17, causal-edge, helix, highheelbgz) to rust-test.yml so this class of regression is caught at PR time. Flagging for the workflow owner rather than bundling a CI change into a test-fix PR.

Generated by Claude Code