feat: add consult plugin, harden council, tighten CI checks

[Droniu]

Summary

bro-code had one second-opinion tool, council, which convenes every model and runs a rebuttal round even when you only want one model's read on the problem in front of you. This PR adds consult for that single-model case, repairs council invocations that could not succeed against current provider CLIs, and widens the CI checks that let those problems ship in the first place.

Context

consult asks one named model — Codex, Grok, or a Claude subagent — and hands it the same access this session has, rather than council's read-only seats. The session's permission mode maps onto each provider: plan stays read-only, acceptEdits confines writes to the workspace, auto auto-approves inside a sandbox, bypass stays bypass. Where the session would stop and ask the user, a headless consultant is denied rather than silently upgraded, since nobody is there to approve. Anything it changes in the working tree is diffed against a pre-run snapshot and reported back, and follow-ups resume the same provider session.

Council's codex path could not succeed against codex-cli 0.153.2: codex exec review rejects -C, refuses --base alongside a positional prompt, and ignores --output-schema entirely, while find-schema.json omitted properties that OpenAI strict structured output requires in required. Review seats now use plain codex exec with the diff target in the brief against a strict-compatible schema, verified with a live call. The grok guidance was wrong in the same way: quota exhaustion was diagnosed as auth (advising a grok login that fixes nothing), --permission-mode plan was accepted and silently ignored, and effort was announced even when the model's catalog reports it unsupported. Seats are now preflighted with a one-turn smoke call before any roster is announced, capped so one runaway seat cannot swallow a run, and gated on file:line evidence before their output counts as a position.

CI listed plugins by hand, so a plugin could be registered in the marketplace and never validated — marketplace --strict passes over a missing source directory and never opens skills. Validation targets now come from marketplace.json, with parity enforced in both directions. The portability scan covered plugins/ only, leaving README, CHANGELOG and workflow content unchecked; it now scans the whole repo.

Test plan

• claude plugin validate .claude-plugin/marketplace.json --strict
• claude plugin validate <each plugin from marketplace.json> --strict — 5/5 pass
• marketplace ↔ plugins/ parity check, both directions
• Portability scan across the whole repo — clean
• markdownlint-cli2 "**/*.md" — 0 issues
• Live codex call against the corrected find-schema.json — valid JSON, no schema rejection
• consult dry runs across the codex, grok and claude paths, plus one live end-to-end consult
• /plugin update on a machine running the published marketplace

https://claude.ai/code/session_01XZZYvBMcU3Q57jF1C4mA7x