NIFAL Audit

Use the Claude Code command as the canonical workflow instead of maintaining a duplicated snapshot.

Load the Canonical Command

1. Read .claude/commands/audit-nifal/SKILL.md completely.
2. Read .claude/commands/_audit-common.md and .claude/commands/_audit-severity.md completely before auditing.
3. Read the live spec docs/engine/nifal.md (plus docs/engine/material-abstraction.md) as the canonical directs, then follow its four tier invariants, risk order, documented-limitation ledger, _audit-validate.sh gate, and Per-Category Tier Matrix report format.

Adapt Claude Conventions to Codex

• Treat the user's text after the skill request as the command arguments described by the canonical skill's argument-hint.
• Resolve each /audit-<name> reference to that audit's SKILL.md under .claude/commands/, and load a sibling audit only at the declared scope seams — parse-side byte errors → nif, mesh-water/env translation → exterior, GpuMaterial layout → renderer, Material save shape → save, BGSM/BGEM/CDB decode → parsers.
• Prefer the repository's codebase-memory graph tools for construction-site caller counts, consumer tracing, and impact analysis; use text search for Option resolve-later fields, raw discriminator fields, per-game if game branches, shader text, docs, and graph gaps.
• Interpret Claude Task-agent instructions as Codex sub-agent delegation only when delegation is available and allowed. Otherwise audit one dimension at a time.
• Treat parked and documented-limitation ledger items as regression guards, not new findings, and where the canonical records known guard holes open the test and confirm a real assertion before trusting a green guard.
• Preserve the canonical output path and do not create GitHub issues; publishing remains a separate audit-publish action.

Do not copy the canonical command body into this skill. The Claude command is the single source of truth.