workflow-fix: Step 5a sibling sync puts main's own files in the branch diff, disabling the #2024 ordering carve-out and false-BLOCKing Step 9c

Goal

Stop the Step 5a sibling-file spec sync (#1972) from converting a PRE-EXISTING, order-dependent test failure on origin/main into a blocking new node at the Step 9c compare. Today the sync writes main's own sibling scripts/issue<M>_*.py / tests/test_issue<M>_*.py files into the branch, which puts them in the branch's three-dot diff; the compare then reads them as branch-touched (paired_skipped: file-in-branch-diff) and the #2024 carve-out — which exists precisely so branch-untouched ordering failures classify ordering_suspect instead of blocking NEW — never fires. The gate blocks a payload that provably cannot cause the failure.

Measured incident (#2296, 2026-08-14)

#2296's payload touches only .claude/skills/issue/SKILL.md, scripts/step9c_baseline.py, scripts/workflow_lint.py, and three test files. Its Step 9c gate returned 2 failed, 9691 passed, 14 skipped in 6396.64s (1:46:36), and the compare classified BOTH failures as blocking new:

new: tests/test_issue2254_reduce.py::test_render_all_on_reduced_tree
     tests/test_issue2254_reduce.py::test_render_all_require_raises_on_skipped
stripped: []
ordering_suspect: []
paired_skipped: [{node: …test_render_all_on_reduced_tree, reason: "file-in-branch-diff"},
                 {node: …test_render_all_require_raises_on_skipped, reason: "file-in-branch-diff"}]
pristine_files_run: ['tests/test_issue2254_reduce.py']
pristine_oracle: scratch-worktree     scratch_sha: 87efce186189…

The failure itself is #2254's, not #2296's:

scripts/issue2254_figures.py:353: in fig_layer_dose_heatmap
    fig.tight_layout()
E   RuntimeError: Colorbar layout of new layout engine not compatible with old
    engine, and a colorbar has been created.  Engine not changed.

Four independent checks establish that the payload is innocent and the red is pre-existing + order-dependent:

1. Byte-identical content across all three trees. git hash-object on the #2296 worktree vs git rev-parse on origin/main and on the oracle sha, for BOTH files: scripts/issue2254_figures.py = 3ed23d34f and tests/test_issue2254_reduce.py = 0a3b6b3bf in all three. The code under test is the same in the failing gate run and the passing oracle run.
2. The tests PASS ALONE in the very worktree whose gate failed — 2 passed in 17.01s for exactly the two blocked node ids, same payload present.
3. The payload has no matplotlib surface at all. git diff origin/main...issue-2296 grepped for matplotlib|pyplot|tight_layout|constrained_layout|colorbar|rcParams returns nothing, so it cannot set a layout engine or create a colorbar.
4. Neither file is in the payload. They entered the worktree solely through the Step 5a sibling-file arm (sync commit 990ec0948d, "sibling-file sync: 17 file(s)").

So a matplotlib layout-engine/colorbar global-state interaction between some earlier test in the 217-file gate set and #2254's figure builder — a condition fully present on origin/main — was attributed to #2296.

Mechanism

• The sibling arm writes main's content into the branch and commits it, so those paths enter git diff <merge-base>...HEAD.
• select_step9c_tests.py sizes the gate from that diff, so the sibling tests are SELECTED (here inflating the gate from the ~61-file invariant set to 217 files and the wall from the documented ~18-min median to 1:46:36).
• The compare's #2024 carve-out keys on branch-UNTOUCHEDness. A synced sibling file is textually branch-touched while being semantically main's own, so the carve-out is skipped and the ordering failure hardens into new.

The three sub-conditions compound: the sync SELECTS the sibling test, then makes it INELIGIBLE for the ordering carve-out, and the larger gate set it produces makes cross-test pollution likelier in the first place.

Proposed fix (direction only — the plan decides)

Make "synced from main by the spec-freshness arm" a first-class provenance signal that the compare can read, so such paths are treated as branch-UNTOUCHED for carve-out purposes while remaining fresh on disk. Candidate shapes:

• Have the sync record the synced path set (a sidecar under .claude/cache/, or a parseable token in its own commit subject — the block already uses the subject shape "sync workflow-surface specs from …" for its dirty-family exclusion), and have _paired_collection_reason / the #2024 ordering classifier subtract that set from "branch-touched".
• Or widen the existing #2206/#2208 revert arm — which already reverts a synced sibling test that fails COLLECTION in the worktree — to cover a synced sibling test that fails at RUN time, using the same restore-branch-era-or-drop disposition.
• Or exclude sibling-synced paths from the selector's diff-derived gate set (they are not the payload under review), which also restores the gate wall to the invariant-set baseline.

Prefer whichever keeps the sync's freshness benefit; do NOT simply disable the sibling arm.

Acceptance

1. A synced-from-main sibling file that fails order-dependently in a gate run classifies ordering_suspect (or is otherwise non-blocking), not new.
2. A genuine branch-authored regression in a file the branch really edited still classifies new — the carve-out must not widen into a hole.
3. The #2296 evidence above is reproduced as a regression fixture (synthetic tree: a sibling-synced path + an order-dependent failure ⇒ non-blocking; the same path branch-edited ⇒ blocking).
4. The gate-set inflation is addressed or explicitly declared out of scope with a reason.

Provenance

Surfaced by the #2296 orchestrator while adjudicating its own Step 9c BLOCK. #2296's own payload (moving the Step 10d mapped-invariant BASELINE leg off the shared repo root) is unrelated to this defect; the two share only the observation that gate baselines/oracles are sensitive to which tree they are cut from. Related: #2293 (the pristine oracle is cut from the root's local HEAD rather than the resolved diff base), #2024/#1832 (the ordering-suspect carve-out this bug disables), #1972 (the sibling-file sync arm), #2206/#2208 (the existing collection-failure revert arm).