fix(task merge): enqueue on a queue-governed branch instead of naming a strategy the queue owns (DIVE-4428)

[lodar]

The defect

5dive task merge shells gh pr merge <url> --squash. main is governed by a GitHub
merge queue, and the queue owns the merge strategy. GitHub answers that combination with

! The merge strategy for main is set by the merge queue
GraphQL: Something went wrong while executing your query

— a 500, which reads as a transient outage and invites a retry that cannot work.

Measured (quinn, 2026-09-13 09:05–09:29Z, DIVE-4428)

Four attempts across 24 minutes, two pull requests, two credentials, one failure mode:

PR	row	attempts	result
#927	DIVE-4412	09:05:13 / 09:05:21 / 09:06:03Z	identical 500 each time
#928	DIVE-4420	09:29:18Z	byte-identical

Both read state=OPEN mergeState=CLEAN mergeable=MERGEABLE, every required check green
(31 of 31 on #928). Not transient. Both rails failed on the same step — entering the queue.

What it cost: #927 was landed at 09:18:48Z by lodar pressing merge by hand; #928 is
still OPEN / CLEAN / MERGEABLE / mergeQueueEntry=null as of this PR. Two verified-good fixes
that the seat which graded them PASS could not land — that is the autonomy number, one step
short of shipping.

The fix

STEP 1 — ask how the branch is governed, in one round trip that also yields what an enqueue
needs. PullRequest.mergeQueue non-null is the governance test. This repo protects main
with a ruleset, which populates no branchProtectionRule, so requiresMergeQueue is the
wrong field to reach for. Verified against #928:

{"id":"PR_kwDOSeJZAM8AAAABDUlPyw","headRefOid":"4924ed95…","isInMergeQueue":false,
 "mergeQueue":{"id":"MQ_kwDOSeJZAM4ABATG","configuration":{"mergeMethod":"SQUASH"}}}

A read that fails leaves every value empty and falls through to the plain merge — a governance
probe must never be the thing that refuses.

STEP 2 — enqueuePullRequest with expectedHeadOid pinned to the graded head. The mutation
returns the queue entry synchronously, so the call that enqueues is the call that proves it;
no second read, no ambiguity. The head pin is the same graded-sha guarantee --match-head-commit
reaches for, enforced server-side: if the head moved, the mutation errors rather than queueing a
tree nobody graded. An isInMergeQueue=true PR is not re-enqueued.

STEP 3 — no queue on the base branch: the ordinary squash merge, unchanged.

Two corrections in the same six lines

• An enqueue is reported as an enqueue. The rail printed <pr> merged by <seat> and
  OK — <ident> merged … is on the target branch over a request that had only been accepted.
  A queue lands it or ejects it — and on DIVE-4154 it ejected 8 minutes later while the rail's
  output said "merged"
  (community/wiki/task-merge-reports-the-enqueue-as-a-landed-merge.md). --json now carries
  merged and enqueued as separate facts.
• GitHub's own message is captured and reprinted, verbatim and first, instead of the refusal
  text pointing at "gh's message above" — output the caller captures into a variable and may never
  have shown.

Evidence

• tests/verifier_merge_standing_unit.sh — 38 passed, 0 failed (8 new arms, section 5: enqueue
  reached, head pinned, ruleset-safe governance probe, already-queued short circuit, disposition
  emitted, caller branches on it, --json carries it, no "read gh's message above").
• tests/task_merge_disposition_unit.sh — 103 passed, 0 failed (unchanged, no regression).
• shellcheck -S error src/task/delivery.sh — clean but for the pre-existing SC2148 (sourced file).

The residual I am signing

The mutation is not exercised live in this PR. A real enqueue is a write against a real PR, and
the only legitimate target — #928 — is another seat's row sitting behind a human gate; landing it
with this seat's admin credential is not mine to press. The mutation's shape is measured working on
this exact repo (community/wiki/main-is-behind-a-merge-queue-so-gh-pr-merge-only-enqueues.md,
DIVE-4289 / PR #880: enqueuePullRequest → {"state":"QUEUED","position":1} after gh pr merge
enqueued nothing), and STEP 1's query is verified against #928 above. The first live proof is this
PR's own merge, which is exactly the call the fix changes.

🤖 Generated with Claude Code