Contract: a conditional sibling-regression allowance per benchmark

[renmengye]

What

An optional per-benchmark regression block. Today a benchmark's single floor sets both its win bar and the regression it tolerates as a sibling. The block separates the two:

- name: rollout-mem
  direction: min
  min_delta_rel: 0.05         # a memory win
  regression:
    free_rel: 0.005           # up to 0.5% worse: always allowed
    max_rel: 0.5              # 50% worse or more: always refused
    requires_gain_rel: 0.05   # in between: only if the climbed benchmark gained at least 5%

• Relative or absolute units, matching min_delta_rel / min_delta. An omitted free tolerance inherits the benchmark's floor, and the hard cap wins at equality.
• Benchmarks without the block keep the original comparison, so every existing contract gates as before.
• Verdict rows, run reports and PR tables name the rule that decided.

Motivating target: jepa.cpp, where a 27% GN-planner speedup cost 33% more memory, a trade the owner accepts. A single 5% memory floor would refuse it.

Verified

• uv run pytest: 3,138 passed, 12 skipped. Ruff, mypy and gitleaks are clean.
• The serial tier fails locally only because the sandbox denies ps.
• contract_cli validates a contract that uses the block.

Review

• Built by Codex.
• Reviewed by Claude.
• For the owner's review; not to be merged without it.

[github-actions]

Round 1 — reviewed head 7489cf52 — reviewer summarizer:hermes/gpt-5.6-terra over coverage+credentials+deployment+general+lifecycle+prose.

terra
Advisory findings from outerloop — the code owner decides. Reply to disagree; the outerloop:no-review label opts this PR out.

Verdict: no defects found.

No findings were submitted by any review lens. Rejected findings: none.