Performance 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-performance/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 parameter parsing and bench discipline, dimension order with their regression guards, pass-inventory verification, merge phase, and 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 when a dimension hands the contract to it (renderer layout/scratch, NIFAL material boundary).
• Prefer the repository's codebase-memory graph tools for hot-path call tracing and guard-symbol discovery; use text search for allocations, bench scripts, timer brackets, literals, configs, docs, and graph gaps.
• Interpret Claude Task-agent instructions as Codex sub-agent delegation only when delegation is available and allowed. Otherwise execute the dimensions sequentially.
• Keep every cost claim measurement-backed per the canonical bench discipline: cite a timer or an observed-vs-ROADMAP delta inside the noise floor, state confidence on speculative Vulkan findings, and never assert harness byte-stability from memory.
• 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.