Fix intermittent Test_Powders results (header parsing) and other undefined behaviour

[willend]

Free-form text area

Please describe what your PR is adding in terms of features or bugfixes:

Fix intermittent Test_Powders results (header parsing) and other undefined behaviour

Branch runtime-ub-fixes, from main. Runtime and component files only, no code-generator changes. Two commits:

1. 6b2138b fixes header-key matching, plus smaller fixes for undefined behaviour.
2. da367a6 fixes reflectivity-lib (optional header keys, KINEMATIC indices).

Root cause of the intermittent Test_Powders results

• Table_ParseHeader() (read_table-lib) used strcasestr(header, key), which matches the first occurrence of the key anywhere in the header, including inside other words.
• PowderN and Single_crystal convert CIF files with cif2hkl into a tmpnam() file. cif2hkl writes its command line, including that random file name (e.g. /var/tmp/tmp.0.aVcX9q), into the header above the real # Vc 181.085 line.
• When the random part contains "vc" (any case), that rank gets V_0 = 0 and absorbs every particle: one MPI rank's contribution goes missing (e.g. 2.94e8 instead of 3.92e8 with 4 ranks).
• Similarly, "dw" in the name sets a bogus Debye-Waller factor: same events, different weights.
• Each rank draws its own temp names, two per rank in Test_Powders, so failures get more frequent with more ranks. The MPI code itself is fine; MPI just makes this bug show up more often.

Fixes

• read_table-lib: Table_ParseHeader() matches keys as whole words, with word boundaries only where the key itself starts or ends with [A-Za-z0-9_]. Keys like "sigma_a " or "e_min=" match as before.
• reflectivity-lib (McXtrace):
  • COATING files: d is optional and defaults to 0, instead of strtod(NULL). It used to be found by accident in "density" (also giving 0), so all McXtrace mirror/coating instruments crashed with whole-word matching until this fix.
  • Type detection: no strtol(NULL) when a file has no Z.
  • KINEMATIC files: the check looked for a 5th key that is never parsed, so it always failed, and the values were read from shifted indices. Both fixed.
• PowderN, Isotropic_Sqw (McStas and McXtrace) and Powder_process (McStas): the line list is calloc'd, because the multiplicity check can read one entry ahead into uninitialised memory.
• Pol_guide_vmirror: copies at most 6 values into its r..ParToFunc[6] arrays. A longer vector used to overflow (flagged by UBSan in ILL_H5).
• FUNNEL, sort_absorb_last(): *multiplier is now also set on the serial path, and there's no division by zero when every particle is absorbed.

Not changed

• tmpnam() is kept: MSVC has no mkstemp(), and with the parser fix the random names are harmless.

Testing (macOS arm64)

• Test_Powders: 40 of 40 MPI runs identical (4 ranks, fixed seed). Before: about 1 in 10–20 runs differed.
• Header-parsing unit test: temp names containing "Vc"/"dw7": old Vc=0 / DW=7, new correct. Across all 343 shipped data files, the neutron powder/S(q,w)/crystal headers parse identically.
• All 444 McStas and McXtrace examples (default parameters, 1e4 rays, seed 1000): identical detector output to main, and no new failures.
• ILL_H5 under ASan/UBSan: clean (before: out-of-bounds writes in Pol_guide_vmirror).
• Linux and Windows/MSVC are untested.

Related

• The wrong values in ILL_H5's {...} vectors with expressions are a code-generator bug, fixed in Cogen improve comments review #2767. This branch only stops the overflow they caused.
• This may make suppress-cif2hkl-Test_Powders / just-suppress-cif2hkl-Test_Powders unnecessary.

---

Declaration of use of AI-tools

• Please add a checkmark here if you used AI-tools during the work for this contribution
• Furter, please describe how / where and for what the tools were used:

^ was 🤖 Generated with Claude Code and using /ponytail full

---

Development OS / boundary conditions

Please describe what OS you developed and tested your additions on, and if any special dependencies are required:

---

PR Checklist for contributing to McStas/McXtrace

For a coherent and useful contribution to McStas/McXtrace, please fill in relevant parts of the checklist:

• My contribution contains something else

  • Explanation is added in free form text above or below the checklist

---

[willend]

Test results (macOS arm64, conda-forge mccode-dev env)

Test	Before (main)	After (runtime-ub-fixes)
Test_Powders, comp=1, 4 MPI ranks, 1e5 rays, fixed seed	about 1 in 10–20 runs wrong: one rank with V_0=0 gives Sph_mon_I=2.94e8 instead of 3.92e8, or a wrong DW changes I but not N	40 of 40 runs identical, Sph_mon_I=3.92196e+08
Header parsing, cif2hkl output named tmp.0.aVcX9q	Vc=0	Vc=181.085
Header parsing, cif2hkl output named tmp.0.k2dw7Q	DW=7 (bogus)	DW not found (correct)
Header parsing, normal files (Al.laz, Rb_liq_coh.sqw)	Vc=66.4 / sigma_abs=0.38	identical
Neutron powder/S(q,w)/crystal header keys, all 343 shipped data files	–	parsed values identical
ILL_H5 under UBSan	3 out-of-bounds writes in Pol_guide_vmirror (index 6, double[6])	clean
McXtrace mirror/coating instruments (17), with whole-word matching only	crash (strtod(NULL) in reflectivity-lib)	fixed by da367a6fa: results identical to main
All 444 McStas and McXtrace examples (default parameters, 1e4 rays, seed 1000)	baseline	444 of 444 identical detector output, no new failures

How it was tested:

• Before and after: used the same installed generator. Branch A's runtime and component files were put first on the -I search path.
• mcrun -I: keeps only the last -I, so one combined search directory was used.
• Sanitizers: Apple /usr/bin/clang. The conda clang's ASan runtime hangs at startup on this macOS.
• Not tested: Linux, Windows/MSVC.

[willend]

https://github.com/mccode-dev/McCode/actions/runs/37625771546/job/112807005872?pr=2770#step:34:391 is hanging on an issue that should be fixed with #2767 in place (correction of {} parsing). Merge.