Rspack perf CodSpeed goal

Use this for performance work where success is judged by a GitHub PR, required correctness validation, review comments, and CodSpeed benchmark feedback.

The agent owns the loop. After the user gives the target benchmark, threshold, and scope, keep making the next best engineering decision without asking for routine confirmation. Continue until CodSpeed confirms the target and correctness gates are green, a real blocker prevents progress, or the configured round limit is reached.

Invariants

• Treat explicit user targets as binding: benchmark name, required percentage, scope, PR, and maximum rounds.
• Default maximum is 10 optimization rounds unless the user specifies another limit.
• One round means one local optimization attempt plus local correctness validation, review validation, push, CI/CodSpeed validation, review-comment triage, and a continue-or-stop decision.
• When running under a persistent /goal, keep the original target intact across resumes. Do not redefine success around partial progress, a pending CI state, or a narrower benchmark.
• Do not end a non-terminal round without scheduling the required 10-minute CI/CodSpeed follow-up.
• Prefer the GitHub connector/plugin for GitHub operations; use gh only when the connector cannot perform the required action. Avoid the 1Password gh shell plugin unless explicitly requested.
• Preserve behavior unless the user explicitly accepts a semantic change.
• Keep changes scoped to the requested ownership boundary. Expand scope only when the measured bottleneck or correctness fix requires it, and explain the expansion.
• Use local verification only for correctness. Run all CodSpeed performance validation in CI and use the CI/CodSpeed result for the current PR head as the source of truth.
• Use append-only commits for normal optimization rounds. Do not routinely amend previous commits or force-push; use history rewriting only for an explicit user-requested rebase or another unavoidable repository operation.