Add run ceilings and demand check (policy-rev 4, v3.1.0)

[ovhirup]

What

Policy-rev 4 / v3.1.0. Each efficiency Stackwich introduces now comes paired with a limit or a count:

• Loop: state a stop condition and a max run count before starting (fixed-cadence or self-paced); stop after 3 no-change runs.
• Graph: state the shard count and why.
• Harness: size delegations to what the verifier/reviewer can check in one pass; done reports count delegations, forks, and loop runs, with token/cost figures only where the environment shows them.
• advisor: new Demand check: plan line and demand-side blast-radius clause.
• New assets/rebound.md: on-demand playbook, read only for cost/loop/fan-out questions. Not part of the policy block.
• Grok and Codex ports bumped to rev 2 with the same limits; CI asserts policy-rev: 4; evals 1–2 updated, 6–7 added.

The <!-- stackwich:v1 --> marker is unchanged, so installs upgrade in place. Block grows 70 → 77 lines.

Verification

• All 8 validate.yml steps pass locally (run with the pip line removed).
• Content checks (rev/version agreement, port revs, no rebound vocabulary inside blocks, eval count) all match expected.
• Advisor reviewed the full diff: APPROVE.

Before merging

• Source links in assets/rebound.md are unverified (no web access during drafting).
• Merging changes main, which the pending directory submission tracks.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 36a98391a3

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Gate parallel shards before aggregating them

For a large batch that is sharded, this new review ceiling is not preserved: the Graph rule at SKILL.md:117-119 still directs every parallel executor result into a single verifier pass. If each delegation is already sized to consume one pass, combining multiple such delegations gives the verifier an N-pass diff at once, recreating the review overload this revision is intended to prevent. Bound the aggregate work to one pass or gate each shard/wave separately before final aggregation.

Useful? React with 👍 / 👎.