[trissim]

Lands after DL1 and after four library releases.

• python-introspect>=0.3.0,<0.4 is pinned (0.3.0 is on PyPI: family members are stored under a tag). ObjectState, formbridge, pyqt-reactive and zmqruntime declare python-introspect<0.3; each needs a release widening that bound before this branch can be installed.
• The storage part (0.6.3: backends state a file's content hash and size, or why they cannot PolyStore#41) has to be released as PolyStore 0.6.4, which requires arraybridge 0.5.0; OpenHCS main is on arraybridge 0.4.1 until refactor/DL1-device-and-memory-types merges. Pins to set then: polystore>=0.6.4,<0.7, arraybridge>=0.5.0,<0.6, submodule external/PolyStore at the 0.6.4 tag (it points at 5301d35 of the Transport mode fix #41 branch until then; pyproject.toml still says polystore>=0.6.3,<0.7, which is wrong).

State after the third hand-off: stored family members name their class (step A) and the record's owner stores it, with both oversized classes back to main's size and local guardrails at rc=0 (step B). Not started: the recorded-content defects, PolyStore 0.6.4, the dataset-root reader and the acceptance re-run. Exact list: "Third hand-off" in the receipt.

DRAFT: handed off mid-job. The receipt docs/refactor/consolidation-20261009/DL3-output-provenance.md has the design, what is implemented, the acceptance results and the ordered list of what remains.

What this does

Gap 6 of the Cellpose/N2V2 investigation (#1230): an output dataset now says what produced it.

• The compiler builds a RunDeclaration (complete pipeline text with every parameter written, the saved global config, and the resolved configuration with the scope and class ObjectState reported for each value). It travels on the CompiledExecutionBundle.
• When execution starts, the parent process builds the RunRecord: per step the function identity (registry key, distribution and version, imports; source hash and text for a custom function), every parameter with its effective value and source (pipeline, function_default, run_time), assigned devices and memory types, and each declared file with the content hash the storage layer reports (PolyStore file_content, 0.6.3: backends state a file's content hash and size, or why they cannot PolyStore#41).
• The record's owner stores it in openhcs_metadata.json under run_records, in a locked update right after the dataset's metadata is published, with produced_files mapping each saved path to the producer identity a viewer item already carries.
• Reader: RunRecord.for_dataset, RunRecord.for_output_file, RunRecord.rebuild(); agent capability openhcs_read_run_record (MCP tool and read-run-record CLI command).

State

• Tests: tests/unit/test_run_record.py (17), tests/unit/test_run_record_capability.py (7), tests/integration/test_run_record_round_trip.py; round trip gives byte-identical TIFF and CSV outputs.
• Acceptance (CPU, Cellpose 3.1.1.3, 3 BBBC039 fields; N2V2 denoise with a model file): passes, one defect found (a declared path left at its default is recorded as written in the pipeline).
• guardrails_local.sh: rc=1 (ratchet rose: the record is handled as string-indexed dicts; code was added inside PipelineCompiler and OpenHCSMetadataTarget). Not yet fixed.
• Not run: smoke_local.sh, parity.
• Depends on 0.6.3: backends state a file's content hash and size, or why they cannot PolyStore#41 (0.6.3) being merged and released.

DL3 · PR draft · guards 0/1 (guardrails failing) · not ready to merge

🤖 Generated with Claude Code