feat(dashboard): emit machine-readable fleet status.json

[ry-ops]

Why

git-steer's autonomous pipeline (heartbeat → ADR-006 escalation → fix workers) already runs on its own and computes the full fleet CVE picture every run. But it only ships that picture as dashboard HTML. So asking "what's the status of my repos?" from a Claude session forced one of two expensive paths:

• a live fleet rescan through MCP tools (Claude as the per-repo loop), or
• parsing 2–3 JSONL files (rfcs.jsonl, cve-queue.jsonl, cache.json) in-context and aggregating.

Both burn Claude spend for data the heartbeat already produced. This was the bottleneck behind hitting monthly limits just to check status.

What

Emit a single small status.json to gh-pages next to index.html:

• fleet — total repos, repos with open alerts, total open alerts, open-by-severity rollup, fixed CVEs, fix rate, avg MTTR
• escalation — ADR-006 posture: hardStop / inLoop lists (threshold matches the escalation script's default of 3)
• repos — per-repo open-alert counts + severity breakdown + escalation status + consecutive fallouts, sorted worst-first

A status query is now one unauthenticated fetch:

https://ry-ops.github.io/git-steer-state/status.json

Pennies, no rescan. git-steer does the work on its schedule; Claude just reads the scoreboard.

Notes

• Built before the dashboard's queue-merge mutates metrics, so the artifact reflects actual open Dependabot alerts + RFC fixes (the canonical "is my fleet clean" view).
• Includes repos that dropped off the managed list but still carry escalation state, so a hard-stop is never hidden.
• Pure additive change to ci-dashboard.mjs; no change to existing dashboard HTML output.
• Verified the builder logic with mock data (hard-stop incl. dropped repo, in-loop, clear, severity rollup, worst-first sort, fleet totals).

🤖 Generated with Claude Code