build(dynacell): pin cubic 0.9.0a3, measured equivalent to 0.9.0a2

applications/dynacell/pyproject.toml

@@ -67,7 +67,9 @@ optional-dependencies.eval = [
   # -- do not bump it casually. Tag, not a bare commit. v0.9.0a2 rewrites only the
   # MicroSSIM RI-factor fit: chunked float64 reduction (peak 40.5 -> 12.9 GB on
   # 576x640x960 float32) and a ValueError on empty input; measured equivalent.
-  "cubic @ git+https://github.com/alxndrkalinin/cubic.git@v0.9.0a2",
+  # v0.9.0a3 fixes single-image FSC/FRC splitting (binomial path, sampling-floor
+  # warning); production's two-image FSC/FRC is untouched -- measured exact.
+  "cubic @ git+https://github.com/alxndrkalinin/cubic.git@v0.9.0a3",
   "dinov3 @ git+https://github.com/facebookresearch/dinov3",
   "dynaclr",
   "hydra-core>=1.2",
@@ -104,7 +106,7 @@ optional-dependencies.eval_gpu = [
 # selection->waveorder, segmentation->cellpose (cellpose imported lazily).
 optional-dependencies.preprocess = [
   "cellpose>=4.2",
-  "cubic @ git+https://github.com/alxndrkalinin/cubic.git@v0.9.0a2",
+  "cubic @ git+https://github.com/alxndrkalinin/cubic.git@v0.9.0a3",
   "iohub>=0.3.6",
   "tqdm",
   "waveorder",

applications/dynacell/src/dynacell/evaluation/CLAUDE.md

@@ -5,7 +5,7 @@ computing) for any GPU-accelerated numerical work — image preprocessing
 before / after model inference, metric calculations, cropping/resizing,
 percentile clips, Gaussian filters, etc. Cubic is a hard runtime
 dependency of the eval extras (`applications/dynacell/pyproject.toml`
-pins `cubic @ git+…@v0.9.0a2`). Do not gate cubic imports behind `try/except`
+pins `cubic @ git+…@v0.9.0a3`). Do not gate cubic imports behind `try/except`
 or fall back to scipy/skimage paths.
 
 **The cubic version is part of the numeric contract, not just a dep.** Metric
@@ -25,8 +25,10 @@ values move across pins — measured 0.8.0a2 → 0.9.0a1: `Z_FSC_Resolution`
   which treats an untagged entry as unconstrained.
 - A bump measured to move no value beyond a stated tolerance lists the old
   version under the new one in `CUBIC_VERSIONS_EQUIVALENT_TO` so its caches are still reused
-  (0.9.0a1 under 0.9.0a2). The map is keyed by the declared version, so the
+  (0.9.0a2 and 0.9.0a1 under 0.9.0a3). The map is keyed by the declared version, so the
   next bump reuses nothing until someone measures an equivalence against it.
+  `check_cubic_pin` accepts those equivalent versions too, so evals queued on
+  the previous venv keep running across a measured-equivalent bump.
 
 The GPU-resident Cellpose-SAM entry point is
 `cubic.segmentation.segment_cpsam` (single host→device upload, masks

applications/dynacell/src/dynacell/evaluation/provenance.py

@@ -22,17 +22,18 @@
 
 This module makes that boundary detectable and non-repeatable:
 
-* :func:`check_cubic_pin` fails a run whose environment disagrees with the
-  version the repo declares, instead of silently producing values from a
-  different numeric stack.
+* :func:`check_cubic_pin` fails a run whose environment holds neither the
+  version the repo declares nor one measured equivalent to it, instead of
+  silently producing values from a different numeric stack.
 * :func:`write_metrics_provenance` stamps the versions beside the metrics,
   and :func:`metrics_provenance_matches` lets the final-metrics cache gate
   refuse a cache built by a different ``cubic``.
 
 A version bump measured to move no value beyond a stated tolerance lists the old
 version under the new one in :data:`CUBIC_VERSIONS_EQUIVALENT_TO`, so its caches
-stay reusable instead of forcing every leaf to recompute. 0.9.0a1 -> 0.9.0a2 is such a
-bump: it rewrites only the MicroSSIM RI-factor fit.
+stay reusable instead of forcing every leaf to recompute. 0.9.0a1 -> 0.9.0a2 and
+0.9.0a2 -> 0.9.0a3 are such bumps: the first rewrites only the MicroSSIM RI-factor
+fit, the second only single-image FSC/FRC paths that production does not run.
 
 The same sidecar stamps the content hash of the CP reference the CP feature
 metrics were scored in (:mod:`dynacell.evaluation.cp_reference`). CP KID/FID/
@@ -53,7 +54,7 @@
 #: The ``cubic`` version this repo is built against. Must equal the pin in
 #: ``applications/dynacell/pyproject.toml``; ``provenance_test.py`` asserts
 #: they cannot drift apart.
-REQUIRED_CUBIC_VERSION = "0.9.0a2"
+REQUIRED_CUBIC_VERSION = "0.9.0a3"
 
 #: For each declared ``cubic`` version, the earlier versions whose metric values it
 #: reproduces within a stated tolerance, so their caches stay reusable. Keyed by the
@@ -71,7 +72,20 @@
 #: the versions; the root finder lands on the same alpha. The other
 #: ``pixel_metrics`` / ``mask_metrics`` columns do not use the RI-factor code and
 #: were bit-identical on 1 FOV x 7 timepoints.
-CUBIC_VERSIONS_EQUIVALENT_TO = {"0.9.0a2": frozenset({"0.9.0a1"})}
+#:
+#: 0.9.0a2 -> 0.9.0a3 changes single-image FSC/FRC only: it judges the split
+#: sampling floor per direction, splits binomial counts before isotropic
+#: resampling, and warns when a single-image checkerboard result is finer than
+#: 2.83 x the pixel size. Production never enables the binomial path, and the
+#: two-image FSC/FRC that ``compute_pixel_metrics`` runs is untouched. Tolerance:
+#: exact. Measured 2026-09-30 on the A549-trained ``fnet3d_paper`` ER and nucleus
+#: ``a549__denv`` leaves (FOV 0/0/fov0000, 7 + 10 timepoints, production eval.yaml
+#: and spectral_pcc/base.yaml kwargs, A40): every pixel column, XY/Z FSC,
+#: FRC_Resolution and Spectral_PCC differ by 0 abs / 0 rel. The binomial +
+#: resample_isotropic control moves 1.8-6.1% under the same harness, so the
+#: comparison can see a change. 0.9.0a1 stays listed: it reproduces 0.9.0a2
+#: (entry above), which 0.9.0a3 reproduces exactly.
+CUBIC_VERSIONS_EQUIVALENT_TO = {"0.9.0a3": frozenset({"0.9.0a2", "0.9.0a1"})}
 
 #: Sidecar written next to ``pixel_metrics.csv`` by :func:`write_metrics_provenance`.
 PROVENANCE_FILENAME = "metrics_provenance.json"
@@ -95,19 +109,28 @@ def installed_versions() -> dict[str, str]:
     return {name: version(name) for name in _RECORDED_PACKAGES}
 
 
+def _accepted_cubic_versions() -> frozenset[str]:
+    """Return the declared ``cubic`` and the versions measured to reproduce its values."""
+    return CUBIC_VERSIONS_EQUIVALENT_TO.get(REQUIRED_CUBIC_VERSION, frozenset()) | {REQUIRED_CUBIC_VERSION}
+
+
 def check_cubic_pin() -> None:
-    """Raise when the installed ``cubic`` is not the declared one.
+    """Raise when the installed ``cubic`` is neither the declared one nor measured-equivalent to it.
+
+    An equivalent version is accepted because its values are, by measurement, the
+    declared pin's values: that is what lets eval jobs queued on the previous
+    venv keep running across a measured-equivalent bump.
 
     Raises
     ------
     RuntimeError
-        If the installed ``cubic`` version differs from
-        :data:`REQUIRED_CUBIC_VERSION`. Fails closed on purpose: a mismatched
-        stack writes plausible values under the wrong numeric contract, which
-        is exactly the failure this module exists to prevent.
+        If the installed ``cubic`` version is not :data:`REQUIRED_CUBIC_VERSION`
+        or listed under it in :data:`CUBIC_VERSIONS_EQUIVALENT_TO`. Fails closed
+        on purpose: a mismatched stack writes plausible values under the wrong
+        numeric contract, which is exactly the failure this module exists to prevent.
     """
     installed = version("cubic")
-    if installed != REQUIRED_CUBIC_VERSION:
+    if installed not in _accepted_cubic_versions():
         raise RuntimeError(
             f"cubic {installed!r} is installed but this repo declares "
             f"{REQUIRED_CUBIC_VERSION!r}. Metric values are not comparable across "
@@ -181,20 +204,18 @@ def metrics_provenance_matches(save_dir: Path, *, cp_space_sha256: str | None) -
     Returns
     -------
     bool
-        True when the recorded ``cubic`` version equals the installed one (or is
-        in :data:`CUBIC_VERSIONS_EQUIVALENT_TO` for the declared pin while the installed one is
-        the declared pin) and, if ``cp_space_sha256`` is given, the recorded binding
-        equals it.
+        True when the recorded ``cubic`` version equals the installed one (or both
+        are the declared pin or listed under it in :data:`CUBIC_VERSIONS_EQUIVALENT_TO`)
+        and, if ``cp_space_sha256`` is given, the recorded binding equals it.
     """
     path = save_dir / PROVENANCE_FILENAME
     if not path.is_file():
         return False
     payload = json.loads(path.read_text())
     recorded = payload.get("versions", {}).get("cubic")
     installed = version("cubic")
-    equivalents = CUBIC_VERSIONS_EQUIVALENT_TO.get(REQUIRED_CUBIC_VERSION, frozenset())
-    reproduced = installed == REQUIRED_CUBIC_VERSION and recorded in equivalents
-    if recorded != installed and not reproduced:
+    accepted = _accepted_cubic_versions()
+    if recorded != installed and not (recorded in accepted and installed in accepted):
         return False
     if cp_space_sha256 is None:
         return True

applications/dynacell/src/dynacell/evaluation/provenance_test.py

@@ -154,8 +154,8 @@ def test_equivalence_only_applies_under_the_declared_pin(tmp_path, monkeypatch):
 
 def test_unmeasured_pin_has_no_equivalents(tmp_path, monkeypatch):
     """A newly declared pin with no measured entry reuses no older cache."""
-    monkeypatch.setattr("dynacell.evaluation.provenance.REQUIRED_CUBIC_VERSION", "0.9.0a3")
-    monkeypatch.setattr("dynacell.evaluation.provenance.version", lambda _name: "0.9.0a3")
+    monkeypatch.setattr("dynacell.evaluation.provenance.REQUIRED_CUBIC_VERSION", "0.9.0a99")
+    monkeypatch.setattr("dynacell.evaluation.provenance.version", lambda _name: "0.9.0a99")
     _stamp_cubic(tmp_path, "0.9.0a1", cp_space_sha256=None)
     assert not metrics_provenance_matches(tmp_path, cp_space_sha256=None)
 
@@ -167,6 +167,21 @@ def test_check_cubic_pin_accepts_the_declared_version():
     check_cubic_pin()
 
 
+def test_check_cubic_pin_accepts_an_equivalent_version(monkeypatch):
+    """A venv on a measured-equivalent version keeps working across the bump."""
+    equivalent = next(iter(CUBIC_VERSIONS_EQUIVALENT_TO[REQUIRED_CUBIC_VERSION]))
+    monkeypatch.setattr("dynacell.evaluation.provenance.version", lambda _name: equivalent)
+    check_cubic_pin()
+
+
+def test_equivalent_environment_reuses_the_declared_pins_cache(tmp_path, monkeypatch):
+    """Equivalence is symmetric: an equivalent venv reuses a cache the declared pin wrote."""
+    _stamp_cubic(tmp_path, REQUIRED_CUBIC_VERSION, cp_space_sha256=None)
+    equivalent = next(iter(CUBIC_VERSIONS_EQUIVALENT_TO[REQUIRED_CUBIC_VERSION]))
+    monkeypatch.setattr("dynacell.evaluation.provenance.version", lambda _name: equivalent)
+    assert metrics_provenance_matches(tmp_path, cp_space_sha256=None)
+
+
 def test_check_cubic_pin_rejects_a_mismatch(monkeypatch):
     monkeypatch.setattr("dynacell.evaluation.provenance.version", lambda _name: "0.8.0a2")
     with pytest.raises(RuntimeError, match="not comparable across"):

applications/dynacell/tools/run_eval_direct.slurm

@@ -82,10 +82,11 @@ ml uv
 # the 0.9.0a1 venv existed, so jobs that did not export DYNACELL_EVAL_VENV kept
 # silently producing 0.8.0a2 FSC/FRC/Spectral_PCC values — 62 of the 121
 # surviving eval logs. ``check_cubic_pin`` now fails such a run loudly, but the
-# default still has to be right. ``cpdino-eval-cubic090a2`` is the
-# ``cpdino-eval-cubic090a1`` freeze with only cubic changed; 0.9.0a1 caches stay
-# reusable under it (``provenance.CUBIC_VERSIONS_EQUIVALENT_TO``).
-export UV_PROJECT_ENVIRONMENT="${DYNACELL_EVAL_VENV:-/hpc/mydata/alex.kalinin/VisCy-venvs/cpdino-eval-cubic090a2}"
+# default still has to be right. ``cpdino-eval-cubic090a3`` is the
+# ``cpdino-eval-cubic090a2`` freeze with only cubic changed; 0.9.0a2 and 0.9.0a1
+# are measured equivalent (``provenance.CUBIC_VERSIONS_EQUIVALENT_TO``), so their
+# caches stay reusable and jobs already queued on the 0.9.0a2 venv still run.
+export UV_PROJECT_ENVIRONMENT="${DYNACELL_EVAL_VENV:-/hpc/mydata/alex.kalinin/VisCy-venvs/cpdino-eval-cubic090a3}"
 export HF_HUB_CACHE="${HF_HUB_CACHE:-/hpc/projects/comp.micro/virtual_staining/models/dynacell/evaluation/hf_cache}"
 export PYTHONUNBUFFERED=1