Phase C — route human attention by risk

[alejandro-nexcade]

What (ROADMAP Phase C)

• Batch CI at zero cost (C4): statusCheckRollup + headRefOid ride the existing bounded enrich call; rollup_to_summary shares one classification fn with summarize (parity test fails loudly on drift). Every card now carries CI without per-PR subprocesses.
• Triage ranking v2 (C1): transparent additive attention score — within-bucket signals (CI, size, sensitive paths) with a provable invariant: their total span (0.70) is smaller than the bucket spacing (1.0), so a Reworked review can never sink below an Approved one. attentionReasons puts the WHY on the card.
• Risk chips (C3): CI x/y, size class S/M/L/XL, sensitive-path, tests-touched — plus the >400-changed-lines "consider splitting" nudge (Cisco sizing data, roadmap F6 folded in).
• Fast lane (C2): teal shelf for small + green + low-risk + stack-root actionable PRs. Same action path as every card — it compresses the decision, not the authority (no new approve/merge surface; §9 traced by review).

Review round

Reviewer caught a CRITICAL in the first cut: pinning base_sha to baseRefOid broke the restack fork-point invariant — reverted (base side of the full-file view keeps its documented drift limitation instead). Also fixed: head_sha refresh guard extended to Reworked (local HEAD leads until push is visible), TS/Rust ERROR classification parity, fast lane restricted to stack roots (no fragmented stacks), per-object signal memoization.

Verified

cargo fmt/clippy -D warnings/test --all (436+) · tsc strict · vitest 179 · vite build · reviewer pass + targeted fix round + coordinator spot-check.

🤖 Generated with Claude Code