Generalize the ACA-branded labels in the shared binary assignment ops

[daphnehanse11]

Summary

The generic assign_binary_from_rate / calibrate_binary_assignment / calibrate_binary_assignment_joint_targets handlers in us_runtime/source_runtime.py hard-coded "ACA binary assignment" / "ACA binary calibration" into every error message, even though they are shared manifest primitives. A failure in any non-ACA stage driving them — empirically verified with #334's person-grain medicaid_take_up stage, where dropping person_weight yielded "ACA binary calibration frame is missing required column(s)" — reported itself as an ACA failure. Closes the follow-up from the #334 review.

• Labels are now keyed to the operation kind: "US binary assignment", "US binary calibration", "US joint binary calibration", matching the file's existing "US ratio computation" / "US support clip" naming. The shared helpers (_binary_calibration_value_state, _preserved_true_mask, _first_available_target_table, _group_columns, _domain_mask) take a label keyword so each handler reports its own name. The _ACA_ prefix is dropped from the corresponding parameter-key frozensets. aggregate_person_to_tax_unit keeps its ACA labels — it is genuinely ACA-specific (Marketplace coverage aggregation) and only the ACA stage uses it.
• Draw salts are deliberately unchanged. aca:{output}, calibrate:{variable}, and joint-calibrate:{variable} are byte-identical, and the aca: prefix is now commented as frozen: existing ACA releases are bit-reproducible against it. The exact-seeded-outcome assertions in the existing ACA tests pass unchanged, which locks this in.
• The calibrate handlers' silent stable_tax_unit_draw default is removed: draw is now a required parameter (the assign handler already required it), and the ACA manifest's three calibrate ops state the previous default explicitly — bit-identical behavior. A future person-grain stage that omits the param now fails loudly instead of silently minting positionally-keyed (not identity-stable) draws under a tax-unit column name. _stable_draws gained a docstring stating its tax_unit_id-vs-positional-index keying and that stages at other grains should mint their own identity-keyed draw column.

Interaction with open PR #334 (medicaid_take_up)

• Materialize Medicaid take-up via anchored count-calibration to CMS state snapshots #334 appends a new medicaid_take_up stage to source_stages.json; this PR edits only the existing ACA stage's calibrate ops. The hunks are disjoint, so the two merge cleanly in either order.
• Materialize Medicaid take-up via anchored count-calibration to CMS state snapshots #334's medicaid ops already pass an explicit draw: "stable_person_draw" param and pre-compute the draw column identity-keyed at person grain, so they comply with the stricter required-draw contract as written.

Tests

• New: a person-grain reproduction of the mislabeled error (asserts the message names "US binary calibration" and contains no "ACA"), required-draw errors for both calibrate handlers, and a person-grain calibration happy path with a declared draw column.
• Full packages/populace-build suite passes (884 tests); ruff check/format clean on the changed files.

🤖 Generated with Claude Code

[MaxGhenis]

Approve

Not purely mechanical — the rename rides alongside a real behavioral change, and both are complete for the PR's scope.

Behavioral change (correct and safe): the draw parameter is now required in both calibrate handlers (previously defaulted to stable_tax_unit_draw). That default was a tax-unit-grain assumption that would silently misfire if a person-grain stage reused the handler, so requiring it is the right hardening. Completeness of the migration checks out: all three calibrate_binary_assignment* operations in source_stages.json are ACA (takes_up_aca_if_eligible ×2, selected_marketplace_plan_benchmark_ratio < 1.0), and this PR adds an explicit "draw": "stable_tax_unit_draw" to each — behavior-preserving. assign_binary_from_rate already required draw pre-PR, so no op regresses. The US manifest is the only consumer of these us_-prefixed handlers, so there's no other call site to update.

Rename is complete for the targeted family. All three shared binary-assignment handlers are fully de-branded: constants (_ASSIGN_BINARY_FROM_RATE_PARAMETER_KEYS, _CALIBRATE_BINARY_ASSIGNMENT_PARAMETER_KEYS, _CALIBRATE_BINARY_JOINT_TARGET_PARAMETER_KEYS) and every error label (US binary assignment / US binary calibration / US joint binary calibration). The test_us_binary_calibration_errors_are_not_aca_labeled test pins the de-branding.

The two handlers left ACA-branded are in-scope-correct, not a miss. _ACA_COMPUTE_RATIO_PARAMETER_KEYS and _ACA_AGGREGATE_PERSON_TO_TAX_UNIT_PARAMETER_KEYS remain, but each has exactly one consumer (the ACA stage) and neither is a binary-assignment op. I confirmed #334's medicaid_take_up stage uses only assign_binary_from_rate and calibrate_binary_assignment (both draw: stable_person_draw) — it never touches compute_ratio or aggregate_person_to_tax_unit, so those two staying ACA-branded blocks no reuse. Worth a one-line note only: they're generic operations that would need the same treatment the day a non-ACA stage reuses them.

Reproducibility care is right: the "aca:" salt prefix in assign_binary_from_rate's draw-fallback branch is deliberately frozen (comment: existing ACA releases stay bit-reproducible), and _stable_draws gains a docstring warning that its positional-index fallback isn't identity-stable and other-grain stages should mint their own keyed draw — exactly what the Medicaid person stage does.

Relationship to #334: this PR generalizes the handlers #334 reuses at person grain. #334 passes draw explicitly so it's safe with or without this PR, but this PR is the hardening that makes a silent grain mismatch impossible. Both edit source_stages.json (ACA region here, medicaid region there), so the second to merge may need a trivial rebase.

CI green (test + wheels, 3.13/3.14), MERGEABLE/CLEAN. Not merging (review-only per request).