workflow-fix: .claude/skills/issue/SKILL.md is 813 bytes from its hard FAIL cap — next edit wedges the Step 9c gate fleet-wide

Problem

workflow_lint.py no-flags currently emits:

WARN: .claude/skills/issue/SKILL.md: 982587 bytes — grandfathered; 813 bytes under its cap (983400).

The no-flags workflow_lint run IS the Step 9c test-verdict gate every /issue session must pass before merging. The size ratchet FAILs (not WARNs) once a grandfathered file exceeds its cap. So the next commit that adds even ~1 KB to SKILL.md flips a fleet-wide WARN into a fleet-wide FAIL, and every concurrent session's Step 9c gate goes red on a condition none of them introduced — the #1388 failure shape (two inline-landed lint-red scripts broke the Step 9c gate fleet-wide).

SKILL.md is among the most frequently edited files in the repo: it grew 3,219 bytes during a single ~5-hour window on 2026-08-16 (979,368 -> 982,587) from four independent landings — a3297aab05 (#2146), 6d40c1dff4 (#2323), 4cb68fcbc8 (#2320), 4992ff1933 (#2318). At that observed rate the remaining 813 bytes is well under one day of normal fleet activity.

The sibling .claude/skills/adversarial-planner/SKILL.md is in the same band (73,229 bytes, 521 bytes under its 73,750 cap) and should be assessed in the same pass.

Why this is urgent rather than routine

The failure mode is cross-session collateral: the session that trips the cap is not the session that gets blocked. Whoever lands the next SKILL.md edit will pass their own inline payload lint gate (their file is fine in isolation) and then every OTHER session's Step 9c gate fails on a byte count. Diagnosing that from a red gate is expensive — the failing line names a workflow-surface file the blocked session never touched.

Scope for the planner (deliberately not pre-decided)

Two broad directions, and the choice is a genuine judgment call:

1. Slim the file. SKILL.md has repeatedly been reduced by relocating long-form blocks into .claude/rules/*.md behind a pointer (the pattern used for codex-ensemble-review.md, disk-hygiene.md, pods.md, compute-backends.md, etc.). The relocation must preserve the always-on load-bearing summary + pointer so nothing silently stops loading, and --check-lessons-index must stay green.
2. Raise the grandfathered cap deliberately. The ratchet exists to force exactly this decision rather than let the file drift. If SKILL.md's size is genuinely justified, the cap should move as a recorded decision with a stated new ceiling — not be bumped reactively each time it is hit.

A third option worth pricing: make the ratchet's failure mode less collateral — e.g. attribute a size FAIL to the file's own last-toucher rather than failing every session's gate. That treats the class, not the instance, but is a larger change.

Acceptance

• workflow_lint.py no-flags leaves .claude/skills/issue/SKILL.md with a stated, deliberate margin below its cap (either by slimming or by a recorded cap decision), and the same assessment is applied to .claude/skills/adversarial-planner/SKILL.md.
• If content was relocated: every relocated block retains an always-on summary + pointer in SKILL.md, --check-lessons-index passes, and no rule file is left unreferenced.
• If the cap was raised: the new ceiling and its rationale are recorded in the same commit as the constant change.
• The full no-flags workflow_lint run shows no NEW failures vs baseline.
• Step 9c universe green.

Provenance

Surfaced during #2324's Step 10 completion audit while verifying Acceptance item 5 (workflow_lint.py no-flags shows no NEW failures). The 813-byte headroom is incidental to #2324 — this round's change set contained zero .claude/ files and contributed none of the growth — and did not gate #2324's completion, which merged cleanly as d829f87272.

A dedup scan over tasks/REGISTRY.json for existing SKILL.md size/cap/byte-budget tasks returned zero matches, so this is not a duplicate. The workflow-fix recursion guard was checked and does not apply: #2324's body carries no workflow_fix_target: line and EPM_WORKFLOW_FIX_SESSION was unset.