add ~max=<n> to --review-with for a per-reviewer iteration cap

[atomantic]

Summary

--review-iterations is global and only ever reached the copilot and @<login> loops. The local agents (codex/claude/agy/grok) and ollama had a hardcoded MAX_ITERATIONS=3 with no way to move it — so a slow local model was expensive to include, and there was no way to budget reviewers differently in one run.

A per-entry ~max=<n> suffix now caps that one reviewer's review → fix → re-review cycles:

--review-with claude~max=2,ollama[qwen3.6:35b]~opt~max=1,codex~max=3

entry := ( <agent> [ "[" <model> "]" ] | "@" <login> ) ( "~opt" | "~max=" <n> )*

<n> is a non-negative integer; 0 means loop-until-clean, bounded by the existing 10-iteration guardrail. The suffix chains with ~opt in either order, is not part of the dedup identity (first occurrence's cap wins on collapse), and rides through /do:config saved defaults (stored canonically as ~opt then ~max=<n>).

The capped vs guardrail split

This needed a second change to be usable. The local-agent and ollama loops report guardrail on cap exhaustion, which is inconclusive in the aggregate and blocks the merge gate — so ~max=1 would have blocked nearly every merge.

All four loops now report capped (clean-equivalent, as copilot/@<login> already did) when the cap was user-configured, and keep guardrail for a built-in ceiling. A ceiling you chose is a budget; one you didn't is an unfinished review. Runs that never pass ~max are unaffected — their built-in caps behave exactly as before. Ollama's coverage-gap rule still outranks both: a partially-reviewed diff reports incomplete, never capped.

~max is ignored (with a warning, not an abort) in --review-mode parallel, where each reviewer runs a single review-only pass and there are no cycles to cap — matching how that mode already handles --reviewer-applies and the stop-mode flags.

Scope

• Loops: multi-reviewer-loop (suffix parsing, dedup rules, cap resolution + precedence, dispatch forwarding, parallel warning, Iterations column in the summary); local-agent-review-loop and ollama-review-loop take {MAX_ITERATIONS}/{MAX_EXPLICIT} instead of hardcoding 3; copilot-review-loop and github-reviewer-loop note where their {REVIEW_ITERATIONS} now comes from.
• Commands: the 7 specs that parse --review-with, plus config.md for saved-default validation and round-trip. /do:rpr forwards ~max to the local-agent and Ollama loops it dispatches and documents that it's a no-op on its bespoke Copilot flow (it doesn't accept --review-iterations either).
• Docs: README suffix explainer, examples, and flag table; .changelogs/NEXT.md.

Test plan

• npm test — 174 pass, 0 fail.
• Added a config.test.js round-trip case covering ~max on a bare slug, a bracketed model, an @<login>, ~max=0, and chained after ~opt, asserting it coexists with a saved review-iterations key (storage stays suffix-agnostic; precedence is the review commands' job).
• Spec-only otherwise — no runtime code parses --review-with; the installer copies these files verbatim.