feat: add behavior-preserving performance reference

[64andrewwalker]

Summary

• Add a conditional behavior-preserving optimization reference for performance,
  latency, throughput, runtime, CPU, memory, allocation, binary-size, power,
  energy, bandwidth, and compute-cost goals.
• Separate the working search metric from final acceptance: keep means a
  better exploratory incumbent, not proof that useful behavior survived.
• Require independent known-bad-sensitive guards, held-out/adversarial
  confirmation, comparable provenance, resource boundaries, and rollback.
• Route the reference from the skill router, classic loop, and plan command, then
  distribute it to Claude, Codex/Agents, and OpenCode mirrors.

This is the optimization counterpart of the ownership split in
pawpaw-dev-skills PR #39.

Flow change

before
  improve metric ──> guard green ──> KEEP ──> implied success

after
  working metric ──> exploratory incumbent
                           │
                           v
  isolated change + independent behavior checks + held-out inputs
                           │
                     accept or reject

Changed files

canonical reference + routes       ████████████████████
Codex / Agents generated mirror    ████████████
OpenCode generated mirror          ████████
AGENTS.md principle                ██

Verification

• bash scripts/transform.sh
  • PASS: Claude canonical source distributed to OpenCode and Codex surfaces
• bash tests/test-orchestrator.sh
  • PASS: 154/154
• bash tests/test-regression.sh
  • PASS: 50/50
• bash tests/test-hooks.sh
  • PASS: 107/107
• cmp across the four behavior reference copies
  • PASS: canonical, Agents, Codex plugin, and OpenCode references match
• git diff --check
  • PASS

Risks / follow-ups

• This is a reasoning and acceptance reference, not a semantic validator. A
  green metric or guard still cannot prove that the chosen observations reflect
  real user behavior.
• The change deliberately adds no comparator, schema, or fixed report template;
  real acceptance remains based on inspectable evidence.

[uditgoenka]

Thanks for the thoughtful behavior-preservation direction. This is intentionally out of scope for the v2.2.2 stabilization release, and the current generated Codex artifacts are not merge-ready: and contain , while the plugin mirrors contain ; those resource paths do not exist. The change also needs a path-existence/parity assertion so transformed references cannot pass byte-comparison checks while remaining unresolvable. Please regenerate from the canonical source with valid package-relative paths and add a test that every routed resource exists. Keeping this PR open for a post-v2.2.2 release.

[uditgoenka]

Exact path correction for the review above: (1) .agents/skills/autoresearch/autoresearch.md routes to .claude/skills$autoresearch/references/behavior-preserving-optimization.md; (2) .agents/skills/autoresearch/plan.md routes to the same nonexistent .claude/skills$autoresearch/... path; and (3) plugins/autoresearch/skills/autoresearch/autoresearch.md plus its plan.md mirror route to nonexistent skills/autoresearch$autoresearch/.... Please make these package-relative and add a path-existence assertion in addition to byte parity.