paper: match 0.5.1

[trissim]

Brings paper.md in line with the code at v0.5.1 (6f69ea4, which is also main). No library code changes. Paths below are under src/arraybridge/.

Nothing was executed: the machine was loaded and no environment was created. Every name and call signature in the code examples was checked by reading the source (ArrayLibrary.named libraries.py:292, convert_to :624, ConversionRoute.planned conversion.py:124, ConversionCounters :470 with installed :426 and format_table :545, AcceleratorDevice devices.py:28, decorator lookup __init__.py:54 and decorators.py:486). The examples were not run.

Claim ledger

#	Old claim	Status	What it became	Source
1	Converts between NumPy, CuPy, PyTorch, TensorFlow, JAX, pyclesperanto	true	kept, libraries now cited	libraries.py:1243, 1287, 1365, 1528, 1697, 1830
2	Users declare a function's memory type with a decorator (@cupy)	true	kept; example now imports the decorator and returns one array	decorators.py:419-493
3	"The decorator handles DLPack zero-copy transfers ... falls back to NumPy bridging"	false	The decorator records the declared libraries, adds streams, OOM retry and dtype restoration. It never converts arguments. Conversion is convert_to / convert_memory; the paper now says so	decorators.py:452-472; converters.py:21-45
4	Decorator "manages per-thread CUDA streams"	true for CuPy and PyTorch only	stated for those two, and only for calls whose arguments are on an accelerator	libraries.py:925-955, 958-1008
5	Decorator "detects and recovers from GPU out-of-memory errors"	partly	"retries after the library's caches are freed"; see 12	libraries.py:535-555
6	"preserves dtypes through conversions"	false as worded	dtype restoration applies to a decorated function's output, not to conversions	decorators.py:313-386
7	Per-framework device, OOM exception, cache and stream APIs (bullets)	true where the source uses them; "TensorFlow has no explicit cache API" unverifiable	kept; the TensorFlow clause removed	libraries.py:1297-1300, 1319, 1345, 1376, 1538
8	"single declarative configuration, generating 36 conversion methods (6 x 6) from ~450 lines of configuration"	false	One class per library; lookup, decorators, MemoryType and conversion between every pair are derived from the classes. No count is given	libraries.py:245-306; types.py:110-124; conversion.py:59-177
9	Table: "Unified API (36 paths)"	false	"One call for every library pair"	libraries.py:624
10	Table: "Automatic fallback"	true only for DLPack to host copy	folded into "Record of the mechanism used"; the fallback is described in Design	conversion.py:153-159, 269-289
11	"The architecture uses data-driven metaprogramming"	false	"an abstract family ... a subclass per library with role parents"	libraries.py:789-1236
12	CupyLibrary listing	true (abridged)	kept, one method dropped for length	libraries.py:1287-1362
13	"Each library implements to_numpy(), from_numpy(), from_dlpack(), and convert_to()"	false for from_dlpack (only DLPackImport libraries: NumPy and pyclesperanto have none)	replaced by the role-parent paragraph	libraries.py:1021-1032, 1243, 1830
14	"Adding a seventh framework requires only adding its ArrayLibrary subclass"	true	kept	ADDING_NEW_FRAMEWORKS.md:3-7; types.py:47-124
15	Thread-local streams via threading.local()	true	kept, scoped to CuPy and PyTorch and to accelerator arguments	libraries.py:961, 979-998
16	"model(image) runs on thread-local stream" (the @torch example)	false in general	A host tensor call runs as written; the example was removed	libraries.py:936-941
17	"true parallel GPU execution is possible"	unverifiable (no measurement in the repo)	removed
18	OOM recovery enabled by default; checks exception types and message patterns	true	kept	decorators.py:430; libraries.py:502-518
19	"clears framework-specific caches"	true for CuPy and PyTorch only	says so; the other four declare no cache to free and only gc.collect() runs	libraries.py:520-523, 1010-1018
20	run_with_oom_recovery listing	true (abridged)	removed for length; behaviour is stated in prose	libraries.py:535-555
21	"the pipeline automatically falls back to CPU processing" (twice, plus Research Impact)	false	"The retry runs on the same device. After two failed retries the error is raised; there is no fallback to the CPU"	libraries.py:543-555
22	"scales and clamps back to uint16"; example of 0.0-65535.0 floats returned as uint16	misleading	The float result's minimum and maximum are mapped onto the full range of the integer dtype, so intensities are stretched to 0 to 65535, not rounded in place	libraries.py:80-86, 195-218, 1217-1231
23	Each library's scale_dtype does the scaling	true	kept	libraries.py:660-666
24	OpenHCS uses CuPy for stitching, flatfield, edge detection; PyTorch via the torch decorator	true in OpenHCS today (ashlar_main_gpu.py:859, basic_processor_cupy.py:285, cupy_processor.py:792, n2v2_processor_torch.py:311)	kept; "segmentation" became "denoising", which is the use I could find	OpenHCS repository
25	"thousands of images ... without coordination overhead"	unverifiable	removed
26	AI disclosure: "data-driven configuration"	false	"one class per library"

New statements, with their source: the five routes and their order (conversion.py:180, 214, 244, 306, 341; CHANGELOG 0.5.1 "Conversion"); one predicate plans and performs (candidates :108-121 is used by both planned :137 and run :153); records name the mechanism used and the declined ones (:160-175, :358-372); ConversionCounters (:470); devices as values (devices.py); a host call does not initialise CUDA (libraries.py:936-941, CHANGELOG 0.5.1).

False or misleading: 9 (rows 3, 6, 8, 9, 11, 13, 16, 21, 22) plus the same wording in the AI disclosure (26). Removed as unverifiable: 3 (rows 7 clause, 17, 25).

For the owner

• The title still says "OOM Recovery". The code retries on the same device after freeing caches, twice, then re-raises. I left the title alone; "OOM Retry" would be the exact word.
• Dtype restoration stretches contrast. A float result with values 100.0 to 200.0 returned to uint16 becomes 0 to 65535 (libraries.py:201-208). The old paper's example implied values are kept. The paper now states what the code does; whether the code should do that is a separate question.
• OOM retry frees nothing for TensorFlow, JAX and pyclesperanto beyond gc.collect() (only CudaLibrary overrides cleanup_loaded).
• OpenHCS pins arraybridge>=0.4.1,<0.5 (pyproject.toml:110 on the branch I read), so the Research Impact section describes OpenHCS on 0.4.x while the paper describes 0.5.1.
• paper.bib: the metaclassregistry URL pointed at github.com/trissim/metaclass-registry; changed to github.com/OpenHCSDev/metaclass-registry (confirmed to exist). The six library entries already in the bib are now cited at first mention. No reference was added.
• Length: 1230 words by wc -w (front matter and code included) against 1019 before. The two new mechanisms (routes, records) account for the difference.
• The comparison table's "—" cells are the original table's "no" marks and were kept.

🤖 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.