Route all analysis microdata loading through policyengine.py's managed path

[MaxGhenis]

Standing rule (2026-08-01)

Analysis/population compute always goes through the latest released policyengine (>=5.0.1): pe.us.managed_microsimulation() / pe.uk.managed_microsimulation(), which return a bundle-pinned country-package Microsimulation with the same MicroSeries interface and provenance on sim.policyengine_bundle. Directly-imported country-package Microsimulation is deprecated for analysis (its default dataset can lag the certified bundle) but remains fine for engine development and tests — YAML tests, variable authoring, and household-level Simulation debugging are untouched.

What changed

The catalog rebuild (#61) already routed the core skills through the managed surface; this PR fixes the two remaining offenders, hardens the prose, and adds enforcement:

• targets/claude/agents/dashboard/backend-builder.md — the last file teaching bare Microsimulation() for statewide impacts. Now: image installs policyengine[us]==LATEST (pins exactly-matched country model + certified bundle — verified against the 5.0.1 pins: policyengine-us==1.764.6, policyengine-core==3.30.0), run_statewide uses managed_microsimulation() and returns policyengine_bundle provenance, the snapshot function pre-caches the certified dataset in the image's HF cache, and UK microsim images get a HUGGING_FACE_TOKEN note (private dataset repo).
• targets/claude/commands/new-tool.md — same Modal-template fixes; adds the managed-path rule for tools that need population-scale results.
• skills/policyengine/SKILL.md — states the latest-release floor in Setup, documents sim.policyengine_bundle and the UK twin, hardens the deprecation note, clarifies that managed_microsimulation(reform=...) is how the country-package reform surface is reached in analysis.
• targets/claude/agents/microsim-runner.md — local path pinned to latest (>=5.0.1).
• tests/test_no_stale_references.py — new anti-rot pattern banning from policyengine_{us,uk} import ... Microsimulation across skills/targets/docs, so the deprecated pattern cannot silently return. The three deliberate deprecation notes carry <!-- stale-ok -->.

Verification

• managed_microsimulation mechanics (keyword-only dataset=, kwargs forwarding to the country package, allow_unmanaged gate, policyengine_bundle stamping, hf_hub_download cache materialization) read from the policyengine.py 5.0.1 tag source, per the repo's verify-before-writing rule.
• uv run pytest --ignore=tests/test_skill_examples.py — 69 passed.
• python3 scripts/build_claude_wrapper.py smoke build — OK.

🤖 Generated with Claude Code

Round 2 — adversarial cross-review fixes

A sol (gpt-5.6-sol) adversarial review of the first round found two blockers, both fixed in the second commit:

1. Live bare-country-package scaffolding survived round 1: skills/policyengine-tools/SKILL.md Pattern C installed policyengine-us==X.Y.Z for society-wide Modal workers, and targets/claude/agents/dashboard/dashboard-planner.md emitted policyengine_package: policyengine-us for a microsimulation endpoint (both ship in the dashboard-builder/app-development/complete bundles). Both now install policyengine[us]==X.Y.Z (>=5.0.1) and route population endpoints through pe.us.managed_microsimulation(); the planner's NJ-pattern prose stops teaching a bare policyengine-us pin.
2. The anti-rot guard was too narrow: it only caught a same-line from policyengine_{us,uk} import ... Microsimulation. It now also catches submodule/whitespace variants, module-qualified policyengine_us.Microsimulation(...), and paren-wrapped multiline imports (whole-file pass); scans .ipynb; and a parametrized corpus (9 banned + 8 allowed spellings) pins its behavior. <!-- stale-ok --> semantics are documented (CONTRIBUTING + lint reason) to also cover explicitly-scoped engine-development examples; residual limits (aliased constructors, import *) are documented in the module docstring.

Round-2 nits fixed: new-tool's Pattern C label, backend-builder's simulation.py import-table row, microsim-runner's local-path summary, and the writing skill's installation example all name the top-level package / managed entry point. A repo-wide sweep for policyengine-{us,uk}==, uv add policyengine-{us,uk}, and pip install policyengine-{us,uk} returns nothing.

Additional verification: HUGGING_FACE_TOKEN confirmed against the policyengine 5.0.1 tag (src/policyengine/bundle.py:596 reads it with HF_TOKEN fallback; the provenance manifest reads HUGGING_FACE_TOKEN only — the note names the universally-honored variable). Tests: 86 passed. Wrapper smoke build: OK.