Fix open-ended reforms applying only to their start year in microsimulation

[MaxGhenis]

Fixes #453

What

reform_dict_from_parameter_values serialised open-ended parameter values (end_date=None) as bare "YYYY-MM-DD" period keys. policyengine-core's Reform.from_dict applies a bare date key for a single instant only, so any open-ended reform — including every flat-dict reform, whose compile path produces open-ended ParameterValues — silently reverted to baseline after its start year in US microsimulation runs.

Every period key now uses the "start.stop" range format:

• explicit end_date → serialised verbatim, as before
• open-ended values → stop the day before the next-later value of the same parameter starts (which also makes the documented step-schedule reform shape order-independent), or 2100-12-31 (OPEN_ENDED_REFORM_STOP) when no later value follows

The UK path (simulation_modifier_from_parameter_values) already handled open-ended values correctly via Parameter.update(stop=None) and is unchanged.

Verification

Unit tests: four existing assertions updated from the buggy shape, three new regression tests (open-ended far-future stop; reverse-order step schedule; open-ended + explicit-end mix). Simulation-level suites tests/test_dict_reforms_on_simulation.py, tests/test_us_reform_application.py, and tests/test_household_impact.py pass (41 tests).

End-to-end, the #453 repro through compile_reform_to_policy → build_reform_dict → Reform.from_dict:

{"gov.irs.income.bracket.rates.7": {"2026-01-01": 0.39}}
before: 2026/2027/2030/2035 = [0.39, 0.37, 0.37, 0.37]
after:  dict = {"gov.irs.income.bracket.rates.7": {"2026-01-01.2100-12-31": 0.39}}
        2026/2027/2030/2035 = [0.39, 0.39, 0.39, 0.39]

and the docs' step-schedule example:

{"gov.irs.income.bracket.rates.7": {"2026-01-01": 0.39, "2028-01-01": 0.40}}
after:  dict = {"...rates.7": {"2026-01-01.2027-12-31": 0.39, "2028-01-01.2100-12-31": 0.40}}
        2026/2027/2030/2035 = [0.39, 0.39, 0.40, 0.40]

Real-world magnitude: a 10-year (2026–2035) revenue estimate of a 39% top rate scored +$17.7B in 2026 and exactly $0.0B in each of 2027–2035 before this fix; after, the years grow +$17.7B → +$33.5B (+$251.4B over the window).

Note

uv.lock includes a one-line true-up of the project's own version (4.18.7 → 4.18.10) produced by uv sync — the committed lock was stale relative to pyproject.toml on main.

🤖 Generated with Claude Code