Close translation parity gaps

[jamestjsp]

Summary

• Add Fortran-reference parity coverage for high-risk translated SLICOT routines across AB, IB, MB, SB, SG, and TG families.
• Fix translation mismatches in MB01UY, MB02CU, MB03BD, MB03LF/MB04CD, and MB04QC-related paths.
• Refresh the translation gap audit and mark resolved parity backlog items complete.

Validation

• .venv/bin/pytest tests/python/ -n auto --reruns 2 --only-rerun "worker .* crashed"
• scripts/check_translation_coverage.py --report-only
• scripts/check_translation_smells.py
• Release native benchmark comparison against origin/main with fresh Meson builds: 13/13 benchmarks passed on both; no clear performance degradation requiring a code change.

Notes

• Existing branch benchmark outliers had matching or near-matching minimum timings, consistent with run noise rather than a branch regression.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 7f99cf44a2

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Use packed leading dimensions when recomputing outputs

In the new recompute_outputs path, a0/b0/d0 are stored as packed n x n buffers (orig), but the GEMM calls still pass lda/ldb/ldd as the leading dimensions (for example at a0 on this line). When any leading dimension is greater than n, this makes BLAS read the packed buffer with the wrong stride, which can produce incorrect transformed matrices or out-of-bounds reads/crashes for valid inputs (COMPQ1=COMPQ2=COMPQ3='I' with padded leading dimensions).

Useful? React with 👍 / 👎.

[jamestjsp]

Fixed in 316e8eb by copying the original A/B/D inputs into packed buffers with DLACPY using the caller leading dimensions, then using n as the packed buffer leading dimension in the recompute GEMMs. Added a ctypes regression that calls mb04cd directly with padded lda/ldb/ldd; tests/python/test_mb04cd.py passes locally.