[trissim]

Symptom

test_main[disk-ImageXpress-3d-multiprocessing-direct-none-none] (and every test_main case) failed with create_projection: input must be a 3D array, got 2D in step 6 (second Z-Stack Flattening).

Bisect

First-parent bisect from fc1f946 (good, 2026-09-24): first bad is a3e6a15, merge of PR #394 (shared runtime plumbing). Inside #394 the first observed bad commit is d092a03 "preserve whole-image domains across main-flow handoff". Its parents are hidden by a separate test_main failure (see 3).

Root cause (two defects; the first predates #394)

1. Stitching ran for z001 only. CallableContract.group_scope_inputs treated every artifact input as the owner of the group scope when the callable declared no group-lineage relation. assemble_stack_cpu(image_tiles, positions) therefore got runtime domain ARTIFACT_MANAGED, and select_lifecycle_anchors kept anchors[:1], so only the first z-plane was assembled. This dates from ccfef5f (July). Before Consolidate runtime execution, producer and projection ownership #394 it was hidden because the loader always gave the lone z001 file a leading axis, so create_projection received (1,H,W).
2. The rank of a cohort depended on how many files matched. After d092a03, a one-member cohort whose producer declared no plane axis was passed whole as 2D, while the same member in a cohort of three was stacked on the runtime-slice axis. This exposed defect 1 and also broke Cell Counting on the single stitched site.

Fix

• CallableContract.main_flow_supplies_primary_image is true when the callable takes main-flow pixels as its primary argument and no adapter manages its artifact inputs. Such a callable's group scope is owned by the main flow. Adapter-managed and relation-declared callables are unchanged.
• ImagePayloadStackComposition.keeps_single_member_whole is the one rule for both the load path (from_loaded_images) and the producer stack cache (with_saved_output_context). Only a persisted whole image (volume domain), or a member that carries its declared axis, stays whole. No case is special-cased for Z.
• G3 note: the second change is in aligned_image_payload.py, one predicate used in two places. Please carry it through the payload refactor.

3. Stale test_main check (not a production bug)

#134 (8224363, in #394) publishes result-only directories with an empty raster inventory on purpose. test_main rejected that, so it now accepts a subdirectory declared as a results_dir.

Tests

• New: test_stored_context_input_beside_main_flow_image_keeps_every_anchor, plus a one-member variant of the RGB cohort reload test. The opaque-domain expectations now load a lone member without a plane axis as (1, …) on RUNTIME_SLICE.
• test_opaque_main_flow_image_domain.py and test_function_patterns.py: 119 passed. Across the 7 related unit files, the 9 failures that remain also fail on main.
• Official30 parity: 29/30 equivalent, identical to main 3d6a6d7 (both fail only cp_tutorial_3d_monolayer, "No image serialization format is registered for suffix '.pkl'").
• test_main: 12/48 pass, listed in the PR comment. The remaining failures are not this bug; they are listed below.

Out of scope, found here

• zarr backend: Zarr conversion output has no declared source projection (zarr-OpenHCS, step 0), and the zarr results metadata entry lacks image_files/available_backends/microscope_handler_name.
• napari: Timed out waiting 30000ms for Napari ack on port 5555, and MetadataWriteError: Saved images lack typed produced addresses.
• zmq: port 7777 is fixed and shared with other agents on this machine; the results mixed plates, so they are inconclusive here.

B1 · PR #1181 · guards n/a (bug fix) · production −9 +50 · tests −9 +62

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[trissim]

test_main passing (12/48): disk-{ImageXpress,OperaPhenix}-direct-{none,fiji}, disk-OpenHCS-direct-{none,fiji}, disk-{ImageXpress,OperaPhenix}-zmq-{none,fiji}. All 48 failed on main before this fix.