Build and verify isolated wrapper release candidates

[MaxGhenis]

Keeps the release-candidate tooling from 818c894e and fixes the two defects it
carries, plus a test rewrite that no review examined.

This stack cannot stand on main. Verified against origin/main:
scripts/release_lock.py, scripts/check_release_credentials.py and
scripts/spm_bundle.py do not exist there; scripts/bundle.py exists but has no
--published-spm; and trace_tro.schema.json exists but its pe:emittedIn enum
is ["local", "github-actions", "policyengine-api"], without the
repository-bundle value release_build.py emits. Base is therefore
max/spm-canonical-wrapper-release-20260910, whose head is already 818c894e.

What the reviews did and did not cover

The R1–R3 chain lives at
/Users/maxghenis/spm-rebuild-20260908/rollout/codex-continuation-20260912/.
The author-side records are the six WRAPPER-RELEASE-BUILD* files —
-PREPARATION.md and -SOURCE-INVENTORY.json (R1, staged patch
39c12b24…), -R2-RESPONSE.md and -SOURCE-INVENTORY-R2.json (patch
cb7c8ef0…), -R3-RESPONSE.md and -SOURCE-INVENTORY-R3.json (patch
f8856478…). The independent verdicts sit beside them in
wrapper-build-source-review/: R1 INDEPENDENT-REVIEW.md "not APPROVE",
thirteen findings; R2 INDEPENDENT-REVIEW-R2.md "APPROVE with limits,
conditional on Findings 1–3"; R3 INDEPENDENT-REVIEW-R3.md "APPROVE with
limits", closed by ROOT-REVIEW-R3.json as source_preparation_approved.

Every reviewer worked without a shell and ran nothing. Two consequences show up
as the defects this PR fixes.

• tests/test_spm_bundle_bootstrap.py is named in none of the six files, none
  of the three frozen inventories, and none of the three reviews. Nobody checked
  whether its assertion still matched the workflow the patch rewrote.
• tests/test_spm_household.py was carried by hash, not read.
  WRAPPER-RELEASE-BUILD-R2-RESPONSE.md:35: "tests/test_spm_household.py
  remains SHA256 061cbc1b…. It has been provided unchanged to root's separate
  household verification agent. This lane did not run household calculations
  against the unrelated installed 2.0.0 model." The three independent reviews
  contain zero occurrences of "household".

assert_source_origin came from R1 finding F9
(INDEPENDENT-REVIEW.md:103-110) and was reviewed as a fix, never exercised.

Commits

35981756 — restore the reviewed household test. Returns
tests/test_spm_household.py to its 1b6c001c content, byte-identical (sha256
fe6df0c7…). The 818c894e rewrite (sha256 061cbc1b…) drives the assisted
case through spm_unit_allocated_housing_subsidy and
spm_unit_allocated_tenant_payment. Neither exists in the pinned
policyengine-us 2.0.0 — zero class <name>(Variable) definitions across the
installed model — so dispatch_extra_variables raises ValueError before any
assertion runs. Both are added by the unified country candidate
PolicyEngine/policyengine-us#9467, verified from its file list:
policyengine_us/variables/household/income/spm_unit/spm_unit_allocated_housing_subsidy.py
and …/spm_unit_allocated_tenant_payment.py. The other three names the rewrite
introduces (pre_subsidy_rent, pha_payment_standard,
receives_housing_assistance) do exist in 2.0.0.

The rewrite returns with the country repin — once #9467 publishes and the
pin in pyproject.toml moves, not before. 1b6c001c is the reviewed
post-Ruling-A content, so this revert leaves Ruling A intact.

14e5db47 — follow the publication gate through its new indirection.
818c894e replaced the Publish job's inline
bundle.py check --published-spm --include-tros --strict-tros with
release_build.py publish-check. The test still searched the Publish steps for
the inline command, so its generator matched nothing and next raised
StopIteration instead of asserting. It now finds the step that carries the
gate (push.yaml:164, index 2) and still requires it before
pypa/gh-action-pypi-publish (index 9). publish_check runs that same check
itself (scripts/release_build.py:1381), pinned by
test_publication_checks_existing_strict_gates_before_member_comparison.
NotifyConsumers is untouched and keeps the dependency-light
python -S scripts/bundle.py check --published-spm.

abf75454 — place a namespace package by its portions.
assert_source_origin refused every policyengine* module whose __file__ was
None. policyengine.tax_benefit_models has no __init__.py, so its __file__
is None while its __path__ points inside the checkout, and any interpreter
that has imported a country model carries it. An instrumented probe over three
import phases located it rather than guessing: bare import policyengine gives
0 offenders; after tests/conftest.py, exactly one — that namespace package;
after tests/test_graph/test_extractor.py, two more. That is why
test_source_origin_rejects_previously_imported_other_checkout failed even when
run alone.

module_origin now places a module by __file__ when it has one, otherwise by
every __path__ portion, each of which must resolve inside the prepared source.
A module with neither, an empty __path__, or any portion outside is still
refused — the shape a synthesized stand-in takes.
tests/test_graph/test_extractor.py was the session's second poisoner: it
installs bare stand-ins for policyengine and policyengine.graph at import
time, i.e. during collection, and never removed them. It now restores
sys.modules in a finally and leaves nothing behind at all.

Three new cases cover the acceptance and the three refusals, and both directions
are mutation-checked: reverting to the pre-fix logic fails the acceptance case,
and accepting any file-less module fails all three refusals.

44604506, b8cf1449, d841dae3, 82721388 and e98d280a are PROGRESS.md. The first overwrote the
prior lane's 178-line record; d841dae3 restores it verbatim beneath this
lane's section.

e0b1f4f2 — hold the strict property where session state cannot reach it.
An adversarial pass over the three fixes closed four gaps. A root policyengine
stand-in used to be refused by the file rule before anything else ran; accepting
a __path__ let it reach files("policyengine"), which raises AttributeError
— a class main() does not catch, so a hold would have ended in a traceback.
The root package's own __init__.py is now settled first. The docstring says
what the check does and does not establish: neither __file__ nor __path__ is
proof, since an in-process caller that can write sys.modules can assign either.
A test pins the shape the allowance newly admits instead of leaving it to be
discovered. The subprocess control from the approved harness patch is adopted as
well, and publish-check is now required before the git tag, not only before the
upload.

Conflicts with prior rulings

Two places where this brief and a prior approved ruling disagree. Reported, not
resolved here.

1. The guard change runs against WRAPPER-R3-CI-DIAGNOSIS-20260913.md §1:
   "Keep the production guard strict. Do not delete arbitrary cached modules or
   accept missing origins merely to pass the full suite." A root-approved patch
   — wrapper-ci-harness-repair-20260913/ROOT-REVIEW.json, approved: true,
   parent 818c894e, patch sha256 9ad1e023… — repairs the same failure in the
   harness with a fresh subprocess and leaves assert_source_origin untouched.
   e0b1f4f2 adopts that subprocess control as an additional test, so the strict
   cross-checkout property is pinned either way and switching to the approved
   patch would cost only the namespace branch.
2. §3 says to preserve the household rewrite, not revert it: "preserve this
   assertion and complete the already approved final-country pin transaction".
   The brief instructed the revert, recorded as returning with the repin. The
   evidence justifying its restoration is
   household-integration-20260912/REPORT.md: 84 wrapper controls against the
   authenticated country 2.0.2rc1.

§1's line-208 reading also means the restore is two-part: the rewrite needs both
the two absent variables and the country's assisted ordinary-resource
independence. Re-adding only the variables would still fail.

A fourth blocker these fixes do not touch

gh pr checks 515 shows Nonpublishing wrapper candidate (release) failing in
12s at scripts/check_release_credentials.py — "Release verification requires
an authenticated Hugging Face access token with role read" — before any strict
gate. This PR addresses the four Test (3.x) failures. #515 stays red until the
credential is settled.

Verification

Locked environment, uv sync --frozen, Python 3.14.4.

Check	Result
pytest tests/test_release_build.py tests/test_spm_bundle_bootstrap.py tests/test_graph -q, before	2 failed, 98 passed (exit 1)
same, after	106 passed (exit 0)
pytest tests -q	1205 passed, 9 skipped, 358s (exit 0)
ruff format --check .	exit 0
ruff check .	11 UP038 (exit 1)

The 11 ruff check findings are all present at 818c894e, all UP038, and none
in a file this PR touches; the touched files pass clean. The local lock resolves
ruff 0.12.11, which still has UP038; CI installs 0.16.7, which does not, and
CI's Lint job passes on this tree.

Not verified here

The commit message on abf75454 names tests/conftest.py →
tests/fixtures/us_reform_fixtures.py:12 as the import path to the namespace
package. That is a real path but not the first one: tests/conftest.py:6 →
tests/fixtures/filtering_fixtures.py:7 reaches it earlier in every session.
Both are unrestored module-level imports; the fix covers either.

🤖 Generated with Claude Code