check_laya_train_rules.sh exits 2 on multi-commit PRs in the merge queue under REBASE (MG 39591283990) #5855

Owner: aprender-a3 (opens the PR per W-1) with a non-author sign-off; found by aprender-57, filed by aprender-flow. One-way door: gate content change (solo + quorum + ANDON).

Evidence

• MG run 39591283990, job 119192582150 (x86-main, started 11:36:28Z, failed 12:03:03Z, head 2134c7b): FAIL scripts/check_laya_train_rules.sh [run] at 12:02:37Z, message: "no merge base for origin/main...HEAD and origin/main is not the first parent of the head" (log line 4881-4882). Its [self-test] passed (line 4880).
• PR feat(flow-010): affected_set.sh diff to crates, reverse-dep closure, nextest filter (E5) #5802, 17 commits. The merge queue's merge_method is REBASE since 08:41Z (SQUASH before).

Cause (read from scripts/check_laya_train_rules.sh run_rules, ~lines 297-310)

The script accepts exactly two shapes: a merge base exists, or origin/main is the head's FIRST parent (the old squash-commit shape). Under REBASE a multi-commit PR is replayed as N commits; the head's first parent is the previous replayed commit, not origin/main, and the shallow clone has no merge base, so it exits 2. A single-commit PR still has origin/main as its parent, so it passes.

Fix direction (not yet designed; needs quorum)

Accept a third, defined shape: head descends from origin/main through first-parent only (rebased chain), diff origin/main..HEAD. Do not widen to "any shape". Add self-test rows for the multi-commit chain (accept) and a head unrelated to base (still exit 2), per the case-table rule.

Falsifier

After the fix, a merge_group run for a PR with >=2 commits reaches RESULT PASS|RED instead of exit 2, and the unrelated-head row still exits 2. Until then, single-commit re-cuts work around it.
Predicted effect: removes one cause of REBASE merge-queue kicks (4/4 enqueues kicked since 08:41Z per my 11:15Z count); the kick reasons for the other 3 are not shown to be this check.