Add a robust 'wait for CI fully green' tool for agents

[jorisjonkers-dev-agents]

Problem

Agents repeatedly hand-roll gh run list/view bash poll loops to wait for a PR/branch's CI to finish. These are fragile and waste turns/tokens:

• []|all on an empty job list returns true → the loop exits before jobs even register (false "green").
• Multiple concurrent workflows per push (Validate, CodeQL, Dependency review) — a single-workflow poll misses required checks.
• State churn queued → in_progress → completed and re-queues mid-run confuse naive status==completed checks.
• Long matrix jobs (e.g. 6 system-test shards) blow past a foreground command timeout, so the poll dies with no verdict.
• Flaky jobs (Playwright 30s timeouts) need a bounded auto-rerun-and-recheck, currently done by hand.
• Distinguishing "fully green" vs "some required check failed" vs "still running" is reimplemented every time, inconsistently.

Ask

A first-class tool (council CLI subcommand and/or an MCP tool) — e.g. council ci-wait / wait_for_ci — that, given a --branch/--pr/--sha (+ --repo):

1. Discovers all check runs + workflow runs for the head commit (not one workflow), waits until every required check reaches a terminal state.
2. Returns a single structured verdict: green | failed | timed_out, with the list of failing jobs and a short failing-log excerpt for each.
3. Handles the empty/registering window correctly (don't conclude before checks appear; wait for the checks to be created).
4. Bounded, cache-aware polling cadence (respect the 5-min prompt-cache window; long backoff for slow matrices) with an overall deadline.
5. Optional --rerun-flaky N: re-run failed jobs up to N times and re-evaluate, reporting which passed only on retry (so genuine failures aren't masked but flakes don't block).
6. Optional --required <checks> allowlist so non-blocking checks (e.g. CodeQL) don't gate the verdict.

Why

This is a universal need across every PR an agent opens; a correct shared implementation removes a recurring class of "false green / silent hang / timed-out poll" bugs and saves significant agent time.