Preserve reform-baseline subsampling regression coverage

[MaxGhenis]

Preserve the reform-baseline subsampling regression coverage from #601 after #573 supplied the production fixes on master. This draft ports only the missing tests and the required changelog fragment; no production code changes.

The examples cover parametric and structural reform results calculated before subsampling, a changed default read without a period input, an override deleted before a reform calculation, and dataset overrides of formula-backed variables preserved in both arms. The deleted-input test drops its obsolete stale-registry assertion and describes the behavior master implements.

The Hypothesis property uses master's recorded-input export and varied populations, weights, formula overrides, reforms, sample sizes, seeds, and combined calculation histories on either arm. It checks export stability, history-independent rebuilding/results, baseline fidelity against an independently subsampled unreformed simulation, and preservation of each retained person's original values. An explicit mixed reform/baseline history example exercises both arms together.

The incompatible structural-ID freezing test and already-merged input-only tests are excluded. #601 remains open for the hub to handle.

Validation against unchanged master production code:

• uv run pytest -q tests/core/test_subsample_computed_values.py — 7 passed (four reform-result cases, changed default, deleted override, and dataset override in both arms).
• uv run pytest -q tests/core/test_subsample_computed_values_property.py — 1 passed (configured for 200 generated examples plus six explicit examples).
• Scoped Ruff format and lint on both new modules — passed.
• git diff --check — passed.

All ported tests pass with production code unchanged from master at 42556d5dff8cde292b657cd1426b5e1351e4c5c8. No defects found and no xfails added.

Full suites and microsimulations were not run, per the hub's single-file validation scope. Documentation changes are not required for tests and a changelog fragment.

[MaxGhenis]

US + core hub merge audit, core#606 at 6b52174cacea079ffcd7ce7aa2baab18d2a1ef49 (squash): regression tests for subsampling a reform simulation after calculating (salvaged from #601).

• Why: Rebuild a subsampled simulation from its inputs, not its calculated values #573 (merged as 4ba5bed9a5) superseded Rebuild a subsampled simulation from its inputs, not its calculated values #601's production changes. The hub's supersession check found that Rebuild a subsampled simulation from its inputs, not its calculated values #601 uniquely held regression coverage for values calculated under a reform before subsampling. This PR carries that coverage onto master; all ported tests pass on unchanged master.
• Scope: tests and a changelog fragment only. No file under policyengine_core/ changes, so there is no behaviour change and no microsimulation impact to measure.
• Review: independent GPT-6.1 Sol review r1 approves this exact head, with no findings (CORE_SEMANTICS: no). It recomputed the template expectations by hand. Its sandbox could not run tests; CI at this head ran them.
• Gates checked live at merge:
  • gh pr checks exits 0;
  • the latest Pull request run at the head concluded success;
  • MERGEABLE, not a draft, no CHANGES_REQUESTED;
  • head pinned.
• Follow-up: Rebuild a subsampled simulation from its inputs, not its calculated values #601 closes as superseded once this merges.
• --admin: used only because a required approving review is the sole block.