Pin the depletion-cut outputs to the generator's own code, not all of src/

[MaxGhenis]

What this fixes

tests/bridge/test_pe_us_depletion_cut_artifact.py::test__artifact__then_the_recorded_commit_is_this_code came in with #506. It ran git diff --quiet <recorded commit> HEAD -- src scripts, so it failed as soon as any later PR changed anything under src/ or scripts/. #509 hit it on its first CI run after #506 merged: AssertionError: src/ or scripts/ changed since the outputs were generated at 5ac5ab53…. #511, #512 and every future code PR would fail the same way.

The change

The check now pins only the generator's own code:

• the generator script, scripts/pe_us_depletion_cut_sample_households.py;
• every script it imports (today scripts/pe_us_minimum_benefit_sample_households.py);
• the populace_dynamics modules any of those scripts import directly.

generator_sources() reads these from the scripts' import statements, so the set follows the code. Today it is 13 files. None of #509, #511 or #512 touches any of them.

What the check doesn't cover. None of these fail this test, and the docstring lists them:

• a change to a module the scripts don't import by name (for example something bridge/policyengine_us.py itself imports);
• the package __init__.py files Python runs on the way to a dotted import (ss/__init__.py for populace_dynamics.ss.params), which today only re-export names;
• data files that imported code reads (for example data/external/ssa_cola_history.json, read by estimates/parameters.py). The old check over src/ and scripts/ didn't pin these either.

The oracle test, test_pe_us_depletion_cut_oracle.py, still reruns the households against policyengine-us when that interpreter is available.

Invariants, and where they are tested

Twelve new tests, on an invented repository built in tmp_path. In it, the generator imports a helper script, a module via from ... import, a name that isn't a module, and (inside a function) a dotted import populace_dynamics.deep.inner. The helper imports a second script, which imports a module no other file names.

• The pinned set is exactly the named code. Each import form resolves, a name that isn't a module resolves to nothing, and the never-imported unrelated.py and deep/__init__.py stay out.
• With no change since the recorded commit, the generator is unchanged.
• New and edited code the generator never imports leaves it unchanged. That covers a new module, an edited unrelated module, an edited unimported package __init__, and a new script.
• A change to any pinned file counts as changed (eight parametrized cases):
  • the generator;
  • the helper;
  • a direct import and its package __init__;
  • a module only the helper imports;
  • the second-level script and its own import, which is the recursion;
  • the dotted import inside a function.
• On the real repository, the pinned set is sorted and unique, every path exists, and it includes the two scripts and both bridge modules.

Mutation check. I ran four deliberate breaks of the resolver, and the new tests fail under each one:

• recursion removed (pending.append deleted): 2 fail;
• only top-level statements walked: 2 fail;
• from package import submodule not resolved: 4 fail;
• no package __init__ fallback: 2 fail.

The earlier one-level fixture passed under the recursion break, which is why the second script was added.

Tests

• tests/bridge/test_pe_us_depletion_cut_artifact.py: 39 passed (27 before, 12 new).
• Tier counts recounted with pytest --collect-only -m <tier>: unit 5,828 → 5,840; total 11,252. Every tier matches tests/tier_counts.json.
• Black 26.5.1 at -l 79 and ruff check pass on every changed file. These are CI's lint commands.

axiom: n/a: test scope only

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
social-security-model	Ready	Preview	Oct 4, 2026 2:13am UTC

[MaxGhenis]

Merge record:

• Merged as a merge commit, 7a99c38, pinned with --match-head-commit c52f81ed.

• CI: gh pr checks 514 exits 0 on c52f81e. MERGEABLE; not a draft; no CHANGES_REQUESTED.

• Independent review (Opus 5.5 lanes):

  • 20261003-161512-pr514-review-r1: APPROVE on 15a4d00.
  • 20261003-220330-pr514-review-r2: APPROVE on 259b62e, a review of the delta since r1. Nothing blocking.

• c52f81e applies two of r2's nonblocking docstring nits:

  • deep/__init__.py is "never pinned", not "never imported";
  • one docstring is shortened to fit 79 columns.

  It changes 3 lines in the test file and no logic. Locally 39 passed and unit tier is 5,840.

• Mutation check before merge: each of four deliberate breaks of the import resolver makes the new tests fail (recursion removed, top level only, no submodule resolution, no __init__ fallback).

• No outside or published numbers: this is test scope only. It unblocks Register gate_epuf against SSA's Earnings Public-Use File (unlocked, report-only) #509, Register and build the SSI asset-test population run (Registration 19; no results) #511 and Document every result's baseline and make the projection baseline selectable (TR2008, TR2026, CBO 2026) #512.