H2 UK fixture: expose composition-changing person mass through the SPI support channel

[MaxGhenis]

Follow-up to #844, stacked on its branch (retarget to main once #844 merges).

The hermetic H2 fixture's SPI support-channel draw selected every household, and household size was aligned to region, so the expansion left weighted person mass at 282.24 → 282.24 and CI could not see the failure class #844 fixed on licensed data. Child presence now follows household_id % 5, independent of the % 4 region cadence, so the region-stratified prior allocation moves mass across household sizes: person mass 272.7 → 272.43 while household mass stays exactly 226.8.

• tools/graph_uk_spine_fixture.py: the three-line composition change; benunit.csv, person.csv, frs_raw/benunit.tab, frs_raw/child.tab, and the descriptor digests regenerated (135 households, 162 persons; uk_spine.json unchanged).
• test_uk_graph.py::test_spi_support_fixture_changes_person_mass_and_conserves_household_mass: runs the real root and support-channel transforms on the committed fixture and asserts person mass changes at the executor's rtol=1e-9 while household mass is conserved exactly in both the weights and the mass log, so a regeneration cannot silently lose the property.
• With _STRUCTURAL_MASS["spi_support_channel"] temporarily set back to conserve, the fixture-backed driver test now fails with Node 'spi_support_channel' mass='conserve' changed stratum 'default': 272.7 -> 272.43 (captured in the lane report); the production declaration is untouched.

Verified here: regeneration is deterministic (rerunning the generator produces no diff), test_acceptance_h_parity.py (H2 with the engine) and test_uk_graph.py pass, ruff and tools/ci_test_groups.py --verify clean. The oracle identity is machine-local and not pinned.

@juaristi22: #835 regenerates this fixture for the 27-stage graph; after rebasing over this, rerunning tools/graph_uk_spine_fixture.py is enough.

Built by a sol lane and reproduced before push.

🤖 Generated with Claude Code