workflow_lint SLURM_GPU_WIDTH_GUARD_RE does not recognize an inherited-CUDA_VISIBLE_DEVICES parse, so correctly-fixed dispatchers cannot exit the grandfather list

Goal

Teach workflow_lint.py's SLURM_GPU_WIDTH_GUARD_RE to recognize an inherited-CUDA_VISIBLE_DEVICES parse as a valid GPU-allocation guard, so a dispatcher that adopts the correct allocation-derived pattern can ratchet OUT of SLURM_GPU_WIDTH_GRANDFATHER instead of being stuck in it.

Observed (task #1336, 2026-08-12, implementation round v21)

scripts/issue1336_dispatch.sh was fixed to derive GPU width and device pins from the SLURM allocation rather than from nvidia-smi (the incident: on a 1-GPU allocation nvidia-smi --list-gpus | wc -l returned 8 and a literal CUDA_VISIBLE_DEVICES=0 override re-pointed compute from the allocated physical GPU 5 to physical GPU 0, another job's device — full write-up and the general rule are filed as #2250). The realized guard is:

EPS_ALLOC_GPUS=()
if [ -n "${CUDA_VISIBLE_DEVICES-}" ]; then
    IFS=',' read -ra EPS_ALLOC_GPUS <<< "$CUDA_VISIBLE_DEVICES"
    NGPU=${#EPS_ALLOC_GPUS[@]}
else
    NGPU=$( (nvidia-smi --list-gpus 2>/dev/null || true) | wc -l )
fi

with all 10 pin sites resolving through ${EPS_ALLOC_GPUS[...]}. This is exactly the pattern the lint's grandfather list exists to push dispatchers toward — but SLURM_GPU_WIDTH_GUARD_RE does not match it, so the file cannot be removed from SLURM_GPU_WIDTH_GRANDFATHER. Surfaced by the implementer as a prose follow-up in the round-v21 report.

Why this matters beyond one file

A grandfather list whose exit condition cannot be satisfied by the correct fix is worse than no list: it stops distinguishing "not yet fixed" from "fixed in a way the regex does not know about". Every future dispatcher that adopts the allocation-derived pattern lands in the same state, so the list silently stops tracking real coverage and the ratchet stops ratcheting.

Proposed fix

1. Extend SLURM_GPU_WIDTH_GUARD_RE to accept an inherited-CVD parse as a guard. Recognize the load-bearing shape rather than one file's spelling: a CUDA_VISIBLE_DEVICES-conditioned branch that populates an array (read -ra / IFS=',') and derives the count from that array (${#...[@]}). Do not key on the EPS_ALLOC_GPUS identifier specifically — a per-issue variable name would make the check pass for exactly one dispatcher and fail for the next correct one.
2. Remove scripts/issue1336_dispatch.sh from SLURM_GPU_WIDTH_GRANDFATHER in the same change, so the ratchet actually tightens and the new recognizer is proven against a real file.
3. Add a lint unit test with BOTH arms: a fixture using the inherited-CVD parse PASSES, and a fixture deriving width from a bare nvidia-smi --list-gpus | wc -l with literal CUDA_VISIBLE_DEVICES=0 pins still FAILS. The negative arm is the point — a recognizer broadened until everything passes would silently retire the check.
4. Re-run the check against every currently-grandfathered dispatcher and report which ones the widened regex now accepts. Any file that newly passes should be removed from the grandfather list only if it genuinely implements the allocation-derived pattern — verify per file rather than trusting the regex, since a newly-passing file is exactly where a too-broad regex would show up.

Acceptance criteria

• The inherited-CVD guard shape is recognized; scripts/issue1336_dispatch.sh is out of SLURM_GPU_WIDTH_GRANDFATHER and the no-flags run does not regress on it.
• A test pins both arms (correct-pattern PASS, nvidia-smi-derived FAIL).
• uv run python scripts/workflow_lint.py (no flags) no worse than its pre-change baseline (~15 pre-existing failures on main — do not chase them; assert only that none newly names an edited file).
• The per-file audit from step 4 is reported in the task body, with a stated verdict for each file that newly passes.

Non-goals

• No change to scripts/issue1336_dispatch.sh (already fixed in #1336 round v21, commit 6ff22758209d5fd642a3f3642479b57900f9c620).
• No new rule prose — the RULE for this trap is #2250's gotchas.md entry; this task is only the mechanical recognizer that lets the ratchet track it.
• No broadening that would let a bare nvidia-smi-derived width pass. If the two cannot be distinguished by regex without false-accepting, say so and narrow the scope to removing the one verified file plus a comment explaining the residual, rather than shipping a recognizer that retires the check.

Provenance

Surfaced as a prose follow-up in #1336's round-v21 implementer report while fixing the underlying GPU-allocation defect that a 1-GPU precheck exposed. Sibling: #2250 (the gotchas.md rule for the trap itself). No compute cost.

Grandfather audit (2026-08-12, post-implementation — acceptance criterion 4)

Re-ran check_slurm_gpu_width with the widened recognizer (_slurm_gpu_width_guard_present: legacy SLURM tokens OR inherited-CVD parse) against every SLURM_GPU_WIDTH_GRANDFATHER entry on main's tree. Result: 22 entries — 22 present, 22 still width-derivation-matched, 0 guarded by the new predicate ⇒ no file newly passes; the grandfather list is unchanged this round. (The task body's "remove issue1336_dispatch.sh" criterion is DEFERRED: the fixed dispatcher exists only on issue-1336-fullcorpora — commit 6ff22758 is not an ancestor of origin/main — so removal now would red the no-flags lint on main. The hygiene guard-adopted WARN + the inverse-calibration pin test force the removal at that branch's merge; a coordination marker was posted on #1336.)

basename	present	width hits	guarded by new predicate
issue1310_dispatch.sh	yes	1	no
issue1335_run.sh	yes	1	no
issue1336_dispatch.sh	yes	1	no (main copy unfixed; branch copy guards → True)
issue1345_dispatch.sh	yes	1	no
issue1417_run.sh	yes	1	no
issue1426_sampled_dispatch.sh	yes	1	no
issue1434_dispatch.sh	yes	1	no
issue1689_dispatch.sh	yes	2	no
issue1738_multiturn_launch.sh	yes	1	no
issue1739_nlmap_dispatch.sh	yes	1	no
issue1769_dispatch.sh	yes	1	no
issue1774_dispatch.sh	yes	1	no
issue1775_fu_run.sh	yes	1	no
issue1775_run.sh	yes	1	no
issue1776_dispatch.sh	yes	1	no
issue1776_p3p4_dispatch.sh	yes	1	no
issue1776_swap_dispatch.sh	yes	1	no
issue2094_dispatch.sh	yes	1	no
issue2162_dispatch.sh	yes	1	no
issue779_ffc_n1m_launch.sh	yes	1	no
issue779_ffc_n50k_launch.sh	yes	1	no
issue923_gpu_phase.sh	yes	1	no

Recognizer probe: _slurm_gpu_width_guard_present = True on the branch copy of issue1336_dispatch.sh (at 6ff22758), False on main's copy — the exact realized shape is accepted; bare nvidia-smi width with literal CUDA_VISIBLE_DEVICES=… pins stays FAILING (negative-arm test pins it).