Narrow over-broad scripts/issue*.py rule globs (3 rules, ~23K tokens on every issue script)

Goal

Narrow the paths: frontmatter globs on the three rules that currently fire on scripts/issue*.py, so that a script which does no marker training, no marker measurement, and no uploading stops pulling those rules into context — without losing coverage on the code that genuinely does those things.

The bug

Three .claude/rules/*.md files carry scripts/issue*.py in their paths: frontmatter:

rule	size	why it fires
upload-policy.md	59,919 B	scripts/issue*.py
marker-leakage-measurement.md	17,089 B	scripts/issue*.py
marker-training-recipe.md	14,782 B	scripts/issue*.py

scripts/issue*.py matches all 1,387 issue scripts in the repo. Every one of them — including pure plotters, pure aggregators, and read-only analysis scripts — loads all three rules: 91,790 B ≈ 23K tokens of always-on context per touch.

Measured instance

Touching scripts/issue1336_full_transfer_lattice.py (a matplotlib plotter: no training, no marker, no upload) loads all three. Measured in a live session 2026-08-11: the plotter's total rule load-set was 453,305 B (~113K tokens), of which 91,790 B was these three rules, none of which can apply to a plotter.

This is not worktree-specific — it fires identically at the repo root. The worktree case just doubles it.

Constraint (the reason this is not a trivial narrowing)

These rules exist to catch real mistakes: upload-policy.md guards the persist-by-default contract; the two marker rules guard the lr ≤5e-6 window and the on-policy marker-at-end DV. A narrowing that stops firing on genuine upload / marker code is a safety regression, strictly worse than the token cost. The fix must demonstrate retained coverage, not just a smaller glob.

Suggested directions (not prescriptive — the plan owns the choice)

1. Split the trigger: keep a broad glob for the small always-relevant summary and move the deep mechanics behind a narrower one.
2. Narrow to the scripts that actually train/upload (e.g. scripts/*train*.py, scripts/*upload*.py, scripts/issue*_train*.py) and verify against the real set of marker/upload-bearing issue scripts.
3. Content-keyed triggering if the harness supports it.

Acceptance

• Enumerate the issue scripts that genuinely train markers / measure marker leakage / upload artifacts; show the new globs still match all of them.
• Show the load-set reduction on a representative plotter.
• workflow_lint.py PASS.

Related

• The worktree double-load (a nested .claude/worktrees/<n>/CLAUDE.md loads a second full copy of the always-on stack, ~37K tokens of pure duplicate) is a DISTINCT issue with a different target and a design-level fix; it overlaps the pending /mnt/eps-data bind cutover (#2132). Not in scope here.