task: the close-evidence line named the wrong evidence (DIVE-2184)

[lodar]

The merge gate's accepting branch claims ancestry and a delegated push. Both are false for a squash-merged PR — and DIVE-2102 closed exactly that way (PR #248 → dfd3d306, head 62e1a83 not an ancestor of main).

Both predictions about this were wrong, which is why the mechanism was measured before touching it. I predicted the gate cannot be satisfied by a squash-merge (blocking). olivia predicted it passes when it should block (permissive). Neither: it accepted for a correct reason and described it falsely. The accepting variable is set by _gate_branch_ident_on_main (attribution — a commit on main whose SUBJECT names the ident, DIVE-2120); _gate_branch_ancestry was demoted and accepts nothing.

Measured rather than inferred: removing the attribution stub from the suite's own ancestor scenario makes the gate refuse (rc=5) with the ancestry stub still set.

The new message states what was measured and claims neither route, because attribution cannot distinguish a delegated push from a squash-merged PR — a squash rewrites the sha, so both look identical to a subject scan.

The test was asserting the message, not the mechanism. The pre-existing arm set both an ancestry stub and an attribution stub, then asserted *ANCESTOR* appeared — green because the message lied. That is how the wording survived DIVE-2120. Corrected, plus an arm that fails if the ancestry claim returns.

31/0 on the ancestry suite. task_deliver_merge_gate_unit's 2 reds are pre-existing (identical on pristine origin/main; DIVE-2114).

🤖 Generated with Claude Code

[lodar]

Second instance, caught live 20 minutes after this PR opened — and it is stronger evidence than the DIVE-2102 case this PR was written from, because here the branch does not exist at all.

Closing DIVE-1925 just now emitted:

warn: DIVE-1925: branch 'dive-1925-zero-downtime-deploy' is an ANCESTOR of main in
lodar/5dive-api — its commits are all on main (no PR needed; delegated push, DIVE-2101).
done=merged-to-main satisfied.

Every load-bearing clause is false:

• "is an ANCESTOR of main" — it closed on attribution. The squash commit is 53121d81, subject DIVE-1925: zero-downtime API deploys via systemd socket activation (#32).
• "no PR needed" — it landed via PR cli: token-burn audit — heartbeat fresh-default off + haiku Explore (DIVE-1210) #32, which the subject literally names.
• "delegated push" — nobody pushed to main; it was a squash merge.
• "branch '...' is" — --delete-branch removed it on merge. The message asserts a present-tense property of a ref that no longer exists.

That last one matters for the fix: the message names the branch as the subject of the claim, and the branch is exactly the thing attribution deliberately stopped depending on (DIVE-2120 moved the search to main precisely because a merged-and-deleted branch was byte-identical to one that never existed). So the wording is a leftover from the mechanism that was replaced.

This is also a different repo (lodar/5dive-api, not 5dive-ai/5dive), so it is not specific to one project's merge style — it is every squash-merged PR in every repo we close against.