Summary

Mark LightProfileLinear.pytree_token as runtime-only via Python's standard __getstate__ / __setstate__ protocol, so it doesn't get lossy-int-to-float persisted through PyAutoFit's database round-trip.

Why

The CI Smoke Tests jobs on autolens_workspace_test/main (both 3.12 and 3.13) have been failing with:

TypeError: __hash__ method should return an integer
  File "autogalaxy/abstract_fit.py", line 133, in linear_light_profile_intensity_dict
    linear_light_profile_intensity_dict[light_profile] = float(reconstruction[i])

Root cause: pytree_token (added in 29b16b47 to give a stable identity across jax.jit flatten/unflatten) is set in __init__ to next(itertools.count()) — an int. When a fit is loaded back via FitImagingAgg, PyAutoFit's database deserializer routes every numeric attribute through the Value SQL row (sa.Float column in instance.py:96), so the saved int comes back as a float (0.0 / 1.0 observed in live trace). Python ≥3.12 strictly rejects non-int returns from __hash__.

What this PR does

pytree_token is a process-local counter increment — not meaningful state to persist (copying a number from one process's counter into another's preserves nothing). Adding __getstate__ / __setstate__ is the canonical Python idiom for marking an attribute as ephemeral.

def __getstate__(self):
    return {k: v for k, v in self.__dict__.items() if k != "pytree_token"}

def __setstate__(self, state):
    self.__dict__.update(state)
    if "pytree_token" not in state:
        self.pytree_token = next(LightProfileLinear._pytree_token_counter)

PyAutoFit's Instance._from_object (at instance.py:73) and Object.__call__ (at model.py:189) already honour these methods. No PyAutoFit changes required. The JAX path uses register_model's attr_const which reads vars(self) directly and is unaffected by __getstate__.

Why not patch PyAutoFit's database serializer

A library-wide fix in PyAutoFit (e.g. an IntValue SQL table) would preserve every persisted int — but pytree_token is the only known consumer of int-typed persisted state whose __hash__ breaks. Galaxy.__hash__ already does int(self.id), GeometryProfile.__hash__ returns id(self). The conceptually-correct fix lives at the layer where the attribute is defined: this attribute should never have been persisted.

Tests added

5 new unit tests in test_autogalaxy/profiles/light/linear/test_abstract.py:

• test__pytree_token_is_int_and_unique
• test__getstate__omits_pytree_token
• test__setstate__assigns_fresh_pytree_token_when_missing
• test__pickle_roundtrip_preserves_int_hash
• test__setstate__preserves_pytree_token_when_present

All pure-numpy (no JAX imports), per the project's library-test policy.

Verified locally

• pytest test_autogalaxy/ → 845 passed (was 840 + 5 new)
• python autolens_workspace_test/scripts/database/scrape/general.py → exit 0, prints "FitImagingAgg Checked" with no traceback

Out of scope

• The AssertionError failures in autolens_workspace_test/scripts/jax_likelihood_functions/... smoke tests — separate root cause, separate PR.

Test plan

• PyAutoGalaxy CI Tests workflow goes green on this PR
• After merge: re-trigger autolens_workspace_test Smoke Tests on main and confirm TypeError: __hash__ method should return an integer is gone

🤖 Generated with Claude Code