Keep SPM valuation out of program benefits and isolate simulation policy

[MaxGhenis]

SPM housing valuation currently leaks into general household benefits, CBO transfer income and CARE/FERA income. A change to an SPM threshold can therefore change a non-SPM result without changing the household's assistance. Use the modeled housing award for general benefits and CBO income, apply the CPUC housing-income exclusion from its effective date, and keep SPM caps in SPM resource calculations. A dependency guard covers both housing and work/childcare caps.

Independent simulations currently share policy and cache state. Give them their own policy ownership and tracing state, preserve deliberate branch-family sharing when a reform replaces parameters or variables, and reject saved poverty outputs as population inputs. Retain warm caches when reusing a prepared root, clone or variable-only reform wrapper at the same structural start. Detach before applying newly detected structural reforms at a different start so arbitrary callbacks cannot mutate the shared source. Verify the bytes of the documented default dataset after download. Fix the test batch runner so the child exit code, complete YAML discovery and a real deadline determine the result.

For SPM resource valuation, sum modeled housing awards within each household, allocate them to SPM units by member share, and apply each unit's housing cap. Allocate tenant payments only from units with positive awards using the same shares. Preserve multiple actual awards, exclude nonrecipients' hypothetical payments, and require SPM geography/composition only where there is an allocated award. Document the producer's assisted-family and timing assumptions and distinguish the tenant-payment allocation assumption from verified Census parity. Reject both new calculated allocation quantities as dataset inputs.

Declare the observed independent-minor role as a native dataset source input even though household examples have a fallback formula. Keep it disjoint from calculated SPM outputs; the producer must carry the observed boolean rather than synthesize a default. Four source-role controls pass in both Core configurations.

The nonpublishing CI matrix builds distinct rc1 and final candidate wheels from clean PR-head checkouts using the real version-bump and guarded lock-refresh helpers. Match the publishing toolchain, pin the isolated build backend closure, and retain actual wheel and source receipts. This keeps tests that mutate an itemization parameter out of the release build.

Validation:

• Final policy-family tests pass 27 cases on each Core version, plus eight existing ownership controls on Core 3.32.5. The tests include real-root reform wrappers and prove detachment occurs before both direct parameter updates and modifier callbacks.
• The affected policy-family and SPM-system modules pass 69 tests with each of Core 3.30.2 and 3.32.5. New regressions first reproduced clone/nested-clone/reform-wrapper cache rebuilding and stale public parameter-lookup tracing; unchanged and structurally modified roots now take the correct paths.
• The 33 release-lock and build-artifact guard tests pass, including the real uv resolver probe.
• 79 runtime/SPM tests pass with Core 3.30.2 and the selected release dependencies (Core 3.32.5); two behavioral-response end-to-end tests pass.
• 115 benefit, CARE/FERA and reform tests pass in both dependency configurations, including unchanged partner fixtures.
• Independent reviewers approved the policy ownership, benefit boundary and batch-runner changes. The ownership review added six adversarial probes; batch review exercised four real subprocess cases.
• The combined allocation/integration suite passes 108 Python tests and three existing single-unit YAML cases in each dependency configuration. Independent allocation review adds four probes per configuration, covering interleaved and permuted mixed households, actual HUD awards, nonrecipient payments, and single-unit parity; no actionable findings remain.

Fable approved the exact source revision 8f74a5152d1e52c33bd7a7cba6340346bd7fe004 against base 2ba7daa30e0e79a40b0597c123e4df76e09b6101, with no actionable findings. Actual dedicated Ubuntu rc1/final CI wheel artifacts were authenticated, including all installed rc1 members; these prebinding artifacts do not qualify the future dataset.

Draft release hold: the associated Microcosm receipt/immigration producer changes, new dataset build, final default-data binding, exact model/data qualification and topline comparison are still in progress. Existing published data certificates do not certify changed producer outputs. This PR does not claim the corrected stack is live or change dependency pins on downstream services.

[MaxGhenis]

Closing as superseded by #9467 (merged 2026-09-15 as 5115fb8), on the US + core hub's PR sweep.

#9467's source table names this PR's current head, 8f74a5152d, as "carried in part", and says it supersedes "the reduced parts of #9448". It carries:

• the CPUC housing exclusion;
• the household benefits and CBO transfer switch to actual housing_assistance;
• the SPM household-to-unit housing allocation;
• the release-build tooling;
• the auto-routing test coverage.

Its "What was dropped, and why" section explains the remaining hunks: the spm.py and system.py isolation (owned by #9463) and the batch-runner implementation. No unique change from this branch is left unaccounted for. Reopen it if one turns up.