maintenance: floor autonerves at the first release carrying the SMALLDAT stamp

The floor was `autonerves>=2026.8.22.1`, which predates the SMALLDAT regime
stamp shipped by PyAutoNerves#153/#154. An autoarray installed from PyPI
therefore resolved an autonerves whose writer emits no card at all:
`should_simulate` saw an absent stamp, read it as "unknown", and fell back to
the shape heuristic — which cannot see capped interferometer datasets, the case
the stamp exists for. The fix was inert for anyone installing rather than
running from a checkout.

`autonerves 2026.8.23.1` is the first release carrying it. Verified by
unpacking both wheels from PyPI rather than inferring from commit dates:
2026.8.22.1 has neither `stamp_small_datasets_regime` nor
`SMALL_DATASETS_HEADER_KEY`, and still writes the `[""]` header comment;
2026.8.23.1 has both and the `""` fix. No release sits between them, so it is
the first. Both floor reasons are now stated in the pin comment — the
JAX-in-base-dependencies requirement (PyAutoLens#687/#702) stays load-bearing
alongside the stamp.

The rationale on the duplicated SMALLDAT literal in `dataset_util.py` was
rewritten because this bump makes its stated reason false — it claimed the
floor named a pre-stamp release, so an import would hard-fail. The duplication
is kept anyway, for a different reason: a floor constrains dependency
resolution only, and an editable checkout or `--no-deps` install can still put
a pre-stamp autonerves on the path. The literal degrades to "card absent" and
falls through to the shape heuristic; an import would be an ImportError at
module load. Silent-safe beats hard-fail.

Both fallbacks (`should_simulate`'s shape heuristic and
`_is_capped_at_the_current_cap`) are untouched — every dataset already on disk
is unstamped and depends on them.

`test_autoarray/structures/arrays/files/array/output_test/array.fits` is
regenerated as a byproduct: it is test output that the suite rmtrees and
rewrites, and the newly-floored autonerves drops the `/ ['']` comment literal
(PyAutoNerves#155). Verified cosmetic — identical cards, values, data and byte
size; only the comment text differs. Leaving it stale would dirty the tree on
every suite run under the new floor.

Closes #482.

Co-Authored-By: Claude <noreply@anthropic.com>

Author: claude

autoarray/util/dataset_util.py

@@ -7,11 +7,19 @@
 SMALL_DATASETS_PIXEL_SCALES = 0.6
 
 # The FITS header card ``autonerves.fitsable.stamp_small_datasets_regime`` writes
-# on every array the stack outputs. Deliberately duplicated here rather than
-# imported: ``pyproject.toml`` floors autonerves at a release that predates the
-# stamp, so an import would hard-fail against a legitimately-resolved older
-# autonerves. Reading the card by name degrades to "absent" instead, which is
-# exactly the fallback path below. Keep in sync with PyAutoNerves#153.
+# on every array the stack outputs. Still deliberately duplicated here rather
+# than imported, though the original reason has expired: the floor in
+# ``pyproject.toml`` now names a stamped release, so importing the constant
+# would resolve. The reason it stays a literal is what the floor does NOT
+# cover. A floor constrains dependency *resolution* only; an editable checkout,
+# a ``pip install --no-deps``, or a hand-built virtualenv can still put a
+# pre-stamp autonerves on the path. Under this literal such an autonerves
+# yields "card absent" and the reader falls through to the shape heuristic
+# below — the safe direction, and the same path every pre-stamp dataset on disk
+# already takes. Under an import it would be an ``ImportError`` at module load.
+# Trading a silent-safe degradation for a hard failure to delete one string is
+# the wrong way round, so the duplication is kept on purpose.
+# Keep in sync with PyAutoNerves#153.
 SMALL_DATASETS_HEADER_KEY = "SMALLDAT"
 
 

pyproject.toml

@@ -24,10 +24,22 @@ classifiers = [
 ]
 keywords = ["cli"]
 dependencies = [
-    # Floor, not a pin (PyAutoLens#687) — bump to the first release with JAX
-    # in autonerves' base dependencies once it exists (PyAutoLens#702), so
-    # backtracking cannot pair this autoarray with a jax-optional autonerves.
-    "autonerves>=2026.8.22.1",
+    # Floor, not a pin (PyAutoLens#687). Two independent reasons hold it up;
+    # neither may be dropped without checking the other.
+    #
+    # 1. JAX in base dependencies (PyAutoLens#702) — the floor must name a
+    #    release carrying `jax`/`jaxlib` in autonerves' *base* Requires-Dist, so
+    #    backtracking cannot pair this autoarray with a jax-optional autonerves.
+    #    Satisfied since 2026.8.22.1; every later release keeps them there.
+    # 2. The SMALLDAT regime stamp (PyAutoNerves#153/#154) — `should_simulate`
+    #    prefers the header card that autonerves' writer stamps, and 2026.8.23.1
+    #    is the FIRST release whose `fitsable.stamp_small_datasets_regime` emits
+    #    it. Against anything older the card is simply absent, the read side
+    #    degrades to its shape heuristic, and capped *interferometer* datasets —
+    #    shape-identical to full-resolution ones — go undetected. That is the
+    #    case the stamp exists for, so the floor is what makes it reachable for
+    #    an installed autoarray rather than a checkout.
+    "autonerves>=2026.8.23.1",
     "astropy>=5.0",
     "decorator>=4.0.0",
     "dill>=0.3.1.1",