feat(pr-triage): add optional Jev pre-filter wave before parallel review

[griffinwork40]

What

Adds an optional Wave 0.5 Jev pre-filter to /pr-triage. Before the parallel review wave, Jev (TypeSafe AI's judgment engine, via the jev MCP server) reads each PR's diff server-side and answers three narrow questions:

Question	Type	Used for
trivial: docs, comments, formatting, lockfiles or dep versions only?	check, yes at >= 0.85	Light route
ready: finished work with no WIP markers or placeholder TODOs?	check, no at <= 0.2	Deferred route
risk: how sensitive is the touched code (0-3)?	score	Light gate (< 1.0) and review order

Routing:

• Light: trivial and risk < 1.0, so /review --light on sonnet.
• Deferred: GitHub draft or ready = no (never for PRs named in --prs), so not reviewed and listed as SKIP. The operator can reply include <N> at the triage gate.
• Full: everything else, unchanged, dispatched in descending risk order.

Safety properties

• Jev never assigns GREEN or BLOCKED. Buckets still come only from /review, and every merge keeps its gate.
• --auto-merge does not cover Light-routed PRs; each needs individual confirmation.
• Fail open: tool missing, path refused, call error, or per-item error (for example an oversized diff) all fall back to the Full route. The skill never retries by pasting diffs as text, since that would pull them into context.
• Data egress: on by default only for public repos. Private repos need --prefilter, because diffs go to TypeSafe's API. --no-prefilter restores the old behavior exactly.
• Diff files are written to .afk/tmp/pr-triage/ under the working directory (the Jev server only reads below its start directory) and deleted at the end.

Verification

Ran Wave 0.5 step by step against the 3 open PRs on the public griffinwork40/agent-afk repo (jev-1.13.0, one jev_triage call, ~0.36s, 9,149 input tokens):

PR	Content	P(trivial)	P(ready)	Risk	Route
#2280	docs-only fix	0.96	0.95	0.01	Light
#2279	code change	0.01	0.95	0.88	Full
#2217	GitHub draft, CI workflow using provider secrets	0.01	0.23	2.94	Deferred (draft)

#2217 only reached P(ready) = 0.23, just above the 0.2 defer cutoff. That is why the isDraft flag was added as a deterministic defer signal, alongside Jev's judgment. The frontmatter parses as YAML. This is a prompt-only change with no code or tests in this repo.

Not verified

• A full end-to-end /pr-triage session using the new wave (the reviews themselves were not re-run).
• Jev's scores vary between runs by roughly 0.13 near 0.65-0.75. The thresholds were set well away from that band (0.85 and 0.2), and a wrong call only changes review depth, never a merge.

[griffinwork40]

Mirrored in the standalone repo: griffinwork40/pr-triage#2