[alxndrkalinin]

Summary

Moves the dynacell eval stack from cubic 0.9.0a1 to 0.9.0a2. Existing 0.9.0a1 metric caches stay valid.

• Pin. Updates the pyproject git-tag pins, uv.lock (the cubic entry only; uv lock --check passes), REQUIRED_CUBIC_VERSION, and the run_eval_direct.slurm default venv, which is now VisCy-venvs/cpdino-eval-cubic090a2.
• Cache compatibility. Adds CUBIC_VERSIONS_EQUIVALENT_TO = {"0.9.0a2": frozenset({"0.9.0a1"})}. It is keyed by the declared version, so the next bump starts with no equivalents. metrics_provenance_matches accepts caches stamped with the versions listed for REQUIRED_CUBIC_VERSION, but only while that pin is installed. The CP-space binding rules are unchanged, and 0.8.0a2 caches are still refused.
• Scan tool. Adds tools/scan_nan_microssim_caches.py, a read-only scan. See the caveat below.

Release scope

git diff v0.9.0a1..v0.9.0a2 touches cubic/__init__.py (the version), cubic/metrics/microssim/ri_factor.py and that file's tests. Nothing else changes. The RI-factor fit now reduces its objective in bounded chunks, accumulating in float64 where 0.9.0a1 took a float32 .mean(), and pools the per-slice elements into preallocated buffers. Peak memory drops from 40.5 to 12.9 GB.

Two edge cases now raise ValueError: get_global_ri_factor on 0 slices and get_ri_factor on empty elements. In dynacell, _calibrate_microssim returns early on 0 positions or pairs. The one remaining empty case (D=0) already raised ValueError under 0.9.0a1 (np.concatenate([])), and the caller catches ValueError, so behavior is unchanged.

Equivalence (measured)

cpdino-eval-cubic090a2 was built from the cpdino-eval-cubic090a1 uv pip freeze with --no-deps, changing only the cubic line. uv pip freeze diff: cubic only.

Acceptance tolerance: MicroMS3IM relative delta ≤ 1e-6, far below the tables' 2-decimal rounding.

MicroMS3IM at production calibration settings. Leaf: A549-trained fnet3d_paper ER a549__denv, 12 FOVs. Calibration used the default max_pairs=12 and seed 42, giving a 576×640×960 float32 pool. MicroMS3IM was then scored on every (FOV, t). No torch model was loaded; the run was on an A40.

	cubic 0.9.0a1	cubic 0.9.0a2	delta
fitted alpha	17.883591651916504	17.883591651916504	0
MicroMS3IM, 108 (FOV, t), range 0.663–0.899			max abs 0, max rel 0

The fit objective itself does change: on a fixed 640×960 float32 pair it moves about 7e-8 relative between the versions (S(α) 0.97342079878 vs 0.97342073207). The root finder lands on the same alpha anyway, so the comparison can detect a difference in the objective even though the result is identical.

Other columns. A full eval of 1 FOV × 7 timepoints of the same leaf ran with max_pairs=4. It produced bit-identical pixel_metrics (PCC, SSIM, NRMSE, PSNR, SI_*, Spectral_PCC, XY/Z_FSC) and mask_metrics columns. These columns do not call the RI-factor code. CP feature columns were not scored in that run; the release diff contains no CP code.

Caveat: all-NaN MicroMS3IM in old caches

Where the 0.9.0a1 calibration ran out of memory (pre-#502 code), a cache holds MicroMS3IM = NaN for every FOV, and 0.9.0a2 would have produced a value. Reusing such a cache keeps the NaN. tools/scan_nan_microssim_caches.py --root /hpc/projects/virtual_staining/training/dynacell lists them.

Scan after the re-eval jobs 36590711–18 finished (2026-09-29 09:36 PDT, rc=0): 930 sidecars, all stamped 0.9.0a1. 35 are all-NaN, and 0 are unexplained. All 35 are the mito A549 mock data_range case, which predates this bump:

• 33 canonical mito/<model>/<train>/a549__mock dirs: 11 models × {a549, ipsc, joint} train sets, except celldiff_r2 (a549 and joint) and celldiff_r2_iterative (ipsc only);
• 2 legacy a549/evaluations_with_embeddings/eval_{phase,vscyto3d_randinit}_mitochondria_mock dirs.

No cache holds an OOM-caused NaN. The gate itself gets no special case.

Tests

Suite: provenance_test.py, tests/test_evaluation_*.py, test_pipeline_cache.py, test_pixel_metrics_parity.py, tools/scan_nan_microssim_caches_test.py.

• 0.9.0a1 venv on origin/dynacell-models: 264 passed, 1 failed.
• 0.9.0a2 venv on this branch: 274 passed, 1 failed.

The failure in both is the known order-dependent test_load_eval_models_pins_morphem_to_the_configured_revision.

Not in this PR

• The run_eval_direct.slurm GPU --constraint is untouched; fix(dynacell/eval): fail the job on MicroMS3IM calibration OOM instead of scoring NaN #502 raises it. This PR's hunk in that file is only the venv default and its comment.
• examples/hf_demo/requirements.txt still pins cubic==0.9.0a1. That pin is deliberate (684f768), and this PR leaves it unchanged.

🤖 Generated with Claude Code

https://claude.ai/code/session_015kuht3npKtb2o5E6h19xbA

[Copilot]

Copilot review overview

🟢 Approval recommended

The dependency and cache-provenance changes are consistent and tested; only minor documentation corrections remain.

Review effort: Balanced
Findings: 2

Open (2)

• Correct release scope beyond chunking · New
• Include buffer and empty-input behavior changes · New

What changed in this PR

Updates Dynacell’s evaluation stack to cubic 0.9.0a2 while retaining compatible 0.9.0a1 metric caches.

Changes:

• Bumps cubic pins, lockfile entry, and evaluation environment.
• Adds guarded cache compatibility for cubic 0.9.0a1.
• Tests and documents the updated provenance contract.

File	Description
uv.lock	Locks cubic 0.9.0a2.
applications/​dynacell/​tools/​run_eval_direct.slurm	Uses the new evaluation environment.
applications/​dynacell/​src/​dynacell/​evaluation/​provenance.py	Adds version-equivalent cache handling.
applications/​dynacell/​src/​dynacell/​evaluation/​provenance_test.py	Tests compatibility and rejection rules.
applications/​dynacell/​src/​dynacell/​evaluation/​CLAUDE.md	Updates evaluation guidance.
applications/​dynacell/​pyproject.toml	Pins cubic 0.9.0a2.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Copilot review overview

🟢 Approval recommended

The dependency bump, compatibility gate, operational scanner, and associated tests are consistent and complete.

Review effort: Balanced
Findings: None

Resolved since last review (2)

• Include buffer and empty-input behavior changes
• Correct release scope beyond chunking