[trissim]

Surface L5 (replaces C4). Receipt with the census and parity table: docs/refactor/consolidation-20261009/L5-arraybridge-operations.md.

Library PR: OpenHCSDev/ArrayBridge#10 (0.4.0, CI 14/14 green, not merged or tagged). The submodule points at its branch head 6f6f907; the pin is arraybridge>=0.4.0,<0.5. Merge this after the library is released.

What changed

• arraybridge: ArrayOperations is a registered family (one leaf per memory type); ArrayOperations.for_memory(MemoryType.X) returns the leaf. Leaves hold only the primitives that differ (max/min/mean/sum, astype, prepend_axis, linspace, ones, assign, outer, floor, accepts). The string *_OPERATIONS roster is deleted. The TensorFlow leaf now names dtypes through its NumPy dtype, which fixes "Cannot interpret 'tf.uint16'" on every decorated TensorFlow call.
• processors/image_stack_operations.py writes the shared operations once: validation, linear/weight masks, apply_mask, max/mean/min projection + one StackProjection family, spatial binning + one SpatialBin family, crop, equalization and percentile normalization (NumPy/CuPy), composite weights and the weighted composite.
• Each <backend>_processor.py keeps its registered names, signatures and contracts. Each name delegates to the shared operation, except where the algorithm genuinely differs (census ≠ cells: sharpen, tophat, and per-backend normalization, equalization and composites).
• StackProjectionMethod gains MIN. NumpyStackProjectionMethod is kept because saved pipelines import it (decision default).

Deletions and certificates

• derived (by image_stack_operations + ArrayOperations.for_memory; covered by per-backend parity): per-backend copies of create_linear_weight_mask (numpy, cupy, jax, torch, tf), create_weight_mask (all 6), apply_mask (5), max_projection (6), mean_projection (5), create_projection + 6 StackProjection strategy families, 2 SpatialBin families + spatial_bin_2d/3d and crop (numpy, cupy), equalization (numpy, cupy), percentile_utils.py, composite weight validation (numpy, cupy, jax, cle), the weighted composite (cupy, jax), 6 _validate_3d_array, and PyclesperantoMaskDimensionality. arraybridge's *_OPERATIONS singletons and name field: derived from the registry.
• derived (by torch.iinfo and TensorFlow DType.max): the hand-written integer-maximum tables in torch create_composite and stack_equalize_histogram, and in TensorFlow stack_equalize_histogram.
• unused: cupy create_composite logging; numpy min_projection (an unregistered public helper with no reader in the repo; create_projection(MIN) serves it).
• tests-only (parity did not exercise these):
  • dtypes int8/int16/int32/uint32/float64 on any backend; parity used uint16, uint8, float32 and constant stacks.
  • JAX on a GPU (the desktop's JAX CUDA plugin sees no GPU; JAX ran on the CPU).
  • The torch decorated path: it is nondeterministic run to run on main itself. The undecorated torch bodies are exact.
  • The TensorFlow create_composite integer clamp: it takes a list of stacks, and parity passed stacks, so it raised TypeError both before and after.
  • The three OPENHCS_SUBPROCESS_NO_GPU branches (the registry already skips modules of disallowed memory types).
  • _check_pyclesperanto_available.
  • The TensorFlow ≥2.12 version probe (TensorFlow 2.8–2.11 was not exercised).

Verification

• Per-backend numeric parity: every public declaration, decorated and undecorated, before (b987b7b) and after. numpy 236/236 and jax-CPU 140/140 (here and on the desktop); cupy 282/282, pyclesperanto 188/188, torch 70/70 undecorated (desktop RTX 2070 SUPER). All exact (bitwise, same dtype); no tolerance was needed. TensorFlow, undecorated: 62 same and 8 error→value (create_projection and create_weight_mask failed on main through the broken nested decorator). Decorated TensorFlow now equals undecorated. See the receipt table.
• Processing/backend unit tests: 104 files, each in its own process. 80 pass; the 24 that fail on the branch fail on b987b7b with identical per-file counts (environment: missing caches/references, Qt).
• arraybridge suite: 247 passed, 17 skipped.
• test_main disk + direct: 3/3 viewer-less cases pass. The 9 napari/fiji cases fail with "no Qt platform plugin" and fail identically on b987b7b.
• official30 parity: 1 passed (30/30), 741 s.

L5 · PRs OpenHCSDev/ArrayBridge#10 + this · guards 4/4 · production −1913 +692 · tests −199 +184

🤖 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.