fix(queue): unstall batch construction -- porcelain parse + pre-push regeneration

[matt82198]

Fixes the merge-queue stall that grew the board to 38 open PRs while the AesopMergeQueue daemon reported PASSED on every pass.

Why nothing was merging

Three independent causes, two fixed here:

1. dirty_paths off-by-one (fixed). It sliced the porcelain line at a fixed column (line[3:]). git() strips its output, so an unstaged modification arrives as M path, not M path, and the slice ate the first character — tests/CLAUDE.md parsed as ests/CLAUDE.md. Every registered generated file then read as an unregistered edit, so worktree_is_safe() answered working tree is dirty forever and no integration branch could be built at all. Reproduced live against the deployed tree. The status field is now matched, not sliced, which is correct for both the raw and stripped forms.

2. Batch union fails a gate every member passed (fixed). A batch is a tree no member ever tested. When two individually-green members each add a test file, the suite counts in tests/CLAUDE.md are correct on both branches and wrong on their union. The pre-push hook ran the drift gate and fail-closed, so the batch push died with [DRIFT] Test suite count mismatch in tests/CLAUDE.md: Python: CLAUDE.md says 235, actual is 236 and the queue was left holding a branch it could never publish (git_failed row, 2026-08-03T08:00:09Z). regenerate_on_batch now runs the generators on the batch branch before the push and commits ONLY paths in the GENERATED_PATHS registry.

3. #712 poisons every batch (not fixed here; #712 held as queue-rejected). #712 wires import_resolution_check into ci.yml as a hard 'Sibling import guard' step. The checker on main does not yet recognize the sanctioned sys.path idiom and flags tools/merge_queue.py:68-69, failing ci (0) on every integration branch. #712's own green check predates merge_queue.py landing on main, so it reads individually-green while poisoning each batch it joins. #724 fixes the checker — land #724 first, then re-admit #712.

No gate is weakened

Same generator the gate itself calls; the gate still runs on the pushed branch and in CI; a generator that fails, is missing, or writes an unregistered path commits nothing and the push fails closed exactly as before (regenerator_failed / regenerator_overreach exception rows).

Bonus defect

The registry names --fix, not --regenerate: the pre-push hook's own failure text advises a flag verify_test_suite_count.py rejects (unrecognized arguments). A test now invokes each generator's real argument parser rather than trusting the string.

Tests

122 in tests/test_merge_queue.py (8 new), all proven to fail against the pre-fix module:

• OLD parse of stripped form: ['ests/CLAUDE.md'] vs NEW: ['tests/CLAUDE.md']
• OLD has REGENERATORS: False

test_merge_queue.py + test_merge_train.py: 164 passed.

🤖 Generated with Claude Code