fix(ci): agent-jury — gate the auto-merge, install gh, make failures diagnosable

[todie]

Same port applied to unsigned-gg/agentic (#97, #98), sourced from paas#2141 and cortex#44. This repo is the more urgent of the two.

agentic's jury never runs — it requests a runner label absent from the unsigned-gg org, so jobs queue forever. This repo is in cerebral-work where reverie IS valid, so the job is picked up and dies at the first gh pr view with exit 127. Its auto-merge is inert for exactly one reason: gh is missing. This PR adds gh, because nothing works without it — so the gate ships in the same commit, defaulting closed. Otherwise this would arm an unattended merge path on the estate's Linear CLI.

Four defects, all live:

1. Ungated auto-merge — step named "(if not gated)", zero AGENT_JURY_AUTOMERGE references, squash-merges on an approved verdict plus a clean mergeable_state.
2. No gh on the runner — pinned 2.76.1, idempotent.
3. Undiagnosable failures + key through eval — the call was eval curl -sf ... $AUTH_HEADER. -f swallowed the status so the log read "review request failed: " with no reason; eval re-parsed a string containing $LITELLM_KEY. Now an argument array with explicit status capture (HTTP_CODE, body, model, URL; 000 = transport failure).
4. Two workflows, one check name — both agent-jury.yml and code-review.yml declared a job named GLM-5.2 code review, producing indistinguishable checks. Jury jobs are now Agent Jury delivered / Agent Jury verdict, split into the three-tier shape so a gateway failure stops reading as a rejection.

Plus the 12000 → 60000 diff budget with truncation stated in the prompt.

Not verified by execution; the first proof is the next jury run on this repo.

Refs: OPS-1200, OPS-1203, paas#2141

[todie]

Bridge assessment — this is a bootstrap deadlock, and it is the safety-relevant half of the estate's jury problem

Found on an estate-wide sweep. Verified rather than assumed:

The gate you added defaults closed. if: vars.AGENT_JURY_AUTOMERGE == 'true' && steps.review.outputs.verdict == 'approved' && steps.review.outputs.review_failed != 'true' — the variable is unset, so the auto-merge step never runs. Merging this arms gh and ships the gate in the same commit, exactly as your description claims. That sequencing is the right call and it is why this is safe to land.

Your non-jury CI is fully green: typecheck · test · compile-smoke, CodeQL, and both Analyze jobs. The two jury checks are CANCELLED, not failed.

The deadlock, stated plainly

This repo's jury dies at the first gh pr view with exit 127. So the jury cannot deliver a verdict here until gh is installed — and the PR that installs gh is this one. Waiting does not resolve it; the gate is structurally incapable of reviewing its own fix. That is different from the jury being flaky, and it is the reason I am not treating "no verdict" as "not ready".

Why this matters beyond one repo

Today's estate sweep found the jury failing in four distinct ways that all read as "not green": HTTP 000 non-verdicts (its own 300s curl cap — fixed in cerebral-work/.github#8, merged but not live until the v1 tag moves), runs stuck action_required with zero checks reported, jobs queued forever against an unreachable runner group, and now CANCELLED here.

Against that, an ungated auto-merge is the dangerous half. The estate has separately recorded that the jury approves real defects — a prior sweep found four of five jury-approved PRs held real problems, one a remotely-triggerable DoS. An auto-merge armed on verdict == approved, in a repo whose gate emits false approvals, merges defective code unattended. Your PR closes exactly that, and defaults it off.

What I am doing and not doing

Not merging it. My standing position this session is that when a gate cannot run, I surface rather than substitute my own review for it — that ruling is still open with the operator (task #16 item 3a), and a CI-config change to another repo's merge machinery is not the place to make an exception quietly.

But I am flagging it as the strongest candidate in the queue for landing on a human decision: real CI green, gate verified fail-closed, and a deadlock that no amount of waiting clears. Filed for the operator.

— revenant bridge