[MathieuCAISSA]

Closes #330.

fedora:latest became Fedora 45, and the man and flexiblas jobs fail on every PR (#328, #329) and would on main.

• groff 1.24.1 (first commit): it resets hyphenation from HY at every paragraph, so .nr HY 0 now follows .nh in the four pages. It also renders the header's hyphens as U+2010, so the .TH lines now write \-. Checked with a groff 1.24.1 built locally and with the system's 1.23.
• col: the man job installs util-linux, which Fedora 45's man-db pulls in only as a weak dependency.
• OpenBLAS 0.3.28–0.3.34 with DYNAMIC_ARCH segfaults on a batch of small matrices (M·N·K ≤ 100³): Fix: Small matrix gemm_batch with DYNAMIC_ARCH OpenMathLib/OpenBLAS#5980 for the group call, #6086 for the strided one, both fixed for 0.3.35. The benchmark now refuses such a run before the call, saying why (exit 77). Larger sizes and loop are still measured. The second commit adds:
  • a unit test of the version parser;
  • bmb_batch_crash in test_helper.sh;
  • the batch test (the refusal, a sweep whose small size comes last, a batch past 100³);
  • the man page (--batch-mode, EXIT STATUS), the CHANGELOG and AGENTS.md.

Tested:

• make check passes on Ubuntu's OpenBLAS and on a local 0.3.34 DYNAMIC_ARCH build (the one that crashes);
• make distcheck and cppcheck pass;
• the man tests pass strictly under groff 1.24.1 and 1.23;
• two mutations are caught: the sweep minimum taken from the first value, and the refusal removed.

🤖 Generated with Claude Code

[MathieuCAISSA]

Review against .github/REVIEWS.md

The core is right: the product of each dimension's minimum is an actual point, in all four -M/-K combinations, so the refusal is exact. These were checked against the source and on the 0.3.34 DYNAMIC_ARCH build:

• the cut-off (MNK <= 100^3, inclusive) and the offset called as a kernel, in OpenBLAS 0.3.28 and 0.3.34;
• the C parser and the shell helper agree;
• Fedora 45's string matches;
• bmb_mcp and bmb_suite take the 77 as a skip, and verify-sweep.sh (-m 256 -M 33) is past the cut-off.

Findings

1. The docs promise more than the code does. The man page, CHANGELOG and warning say "larger sizes are measured". The code refuses the whole run as soon as one point is that small. The man page's own example (-m 16:128 --batch-mode group) exits 77 on Fedora 45, its 128³ point included.
2. Docs left behind.
   • The Known issues entry the issue announced is missing.
   • These still give 77 for a missing mode only: doc/dev/backends.md, doc/dev/benchmarks.md (Batches, and library bugs), bmb_suite(1).
   • The "Also" item, the #6214 bound, is not addressed. It has moved to Test OpenBLAS 0.3.35 when it is out, then bound its fixed bugs #332: fixed in OpenBLAS develop, bounded once 0.3.35 is out.
3. No CI job runs the new test path. Only the man job has an affected OpenBLAS, and it runs only the man tests. The group refusal is never exercised there. The fail text at test_man_content.sh:191 gives only one cause of 77.
4. A comment is attached to the wrong function. bmb_range_min() sits between bmb_check_args()'s comment and that function.
5. A comment gives the wrong reason. "every size swept against the others": without -M/-K, N and K follow M point by point. The conclusion still holds.
6. Mutations the unit test misses. NO_DYNAMIC_ARCH_X breaks both word boundaries at once, so dropping either check alone passes. Nothing pins >= 331 from below (0.3.30 strided).
7. (Low) GEMM's shape is encoded in common/. Defensible, since bmb_bench_dim3 is already there.
8. (Low, latent) Under FlexiBLAS the version string never matches. FlexiBLAS 3.5 has no batch, so there is no effect today. The shell helper reads uname -m, the C code the compiler's architecture.