Renderer 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-renderer/SKILL.md completely.
2. Read .claude/commands/_audit-common.md and .claude/commands/_audit-severity.md completely before auditing.
3. Follow the canonical command's delta-scoped dimensions, guard-first verification, symbol-anchored evidence rules, per-dimension outputs with merge and dedup, and report format; audit against docs/engine/shader-pipeline.md and docs/engine/memory-budget.md rather than restating them.

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 its declared scope boundary (exterior pass content, the NIFAL material boundary, performance pool policy) instead of duplicating its checklist.
• Prefer the repository's codebase-memory graph tools for symbol discovery, callers, resource lifetimes, and impact analysis; use text search for shader text, #[repr(C)] structs, barrier and flag literals, configs, docs, and graph gaps.
• Interpret Claude Task-agent instructions as Codex sub-agent delegation only when delegation is available and allowed. Otherwise audit the dimensions sequentially in risk order.
• Keep Vulkan claims that cargo test cannot see explicitly labeled needs-RenderDoc (or a BYRO_VALIDATION=1 sync-validation run), as the canonical command requires.
• 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.