gotchas.md: parallel lm-eval fan-out races library nltk download — pre-stage once before forking (from #2203)

Goal

Add a .claude/rules/gotchas.md entry (and, if judged appropriate by the implementer, a matching launcher-template note in the experimenter agent spec) for the parallel lm-eval fan-out nltk staging race, so future multi-GPU capability/eval fan-outs pre-stage nltk resources once before forking.

Context (incident, #2203 Phase 2, 2026-08-09)

A 4-way CVD-pinned lm-eval capability fan-out (one process per GPU, shared container) crashed 2 of 4 shards at IFEval dataset prep: each process, on first use, triggers the lm-eval library's own nltk.download('punkt_tab') into the shared /root/nltk_data; concurrent downloads collide (one process removed punkt_tab.zip mid-unzip of another), producing LookupError: Resource 'punkt_tab' not found in the losers. The surviving shards completed normally, which makes the failure look nondeterministic/flaky rather than structural.

• Failure signature to document: [nltk_data] [Errno 2] No such file or directory: '.../punkt_tab.zip' followed by LookupError: Resource punkt_tab not found, in SOME shards of a parallel fan-out while sibling shards succeed.
• Fix pattern (validated on the #2203 retry launcher): stage the resource ONCE, single-threaded, as a launcher gate BEFORE the fan-out — nltk.data.find('tokenizers/punkt_tab') except LookupError → nltk.download('punkt_tab') → re-find() — and fail the launcher loud on rc!=0. This mirrors the existing stage-once-before-fan-out discipline for HF artifact staging (the #2203 axis pre-stage).
• Scope note: the nltk call lives in library code (lm-eval IFEval utils), not repo scripts, so the fix layer is the launch wrapper / bootstrap, not experiment code.

Acceptance

1. gotchas.md gains the entry (signature + fix pattern + one-line incident cite), with the LESSONS.md index row trigger text updated if the gotchas.md trigger line changes.
2. workflow_lint.py --check-lessons-index passes.