Recover CensusAPI.load() metadata from name + long data

[jinskeep-morpc]

Summary

CensusAPI.load() previously depended on a private _morpc block written into the resource descriptor. This removes that dependency: save()/create_resource() no longer write _morpc, and load() recovers all six constructor arguments from the canonical dataset name and the long-format data.

Recovery map (_recover_metadata(name, long)):

Arg	Source
survey, year	uniform survey / reference_period columns
variables	base codes re-suffixed from the value-type columns (new _recover_variable_codes); only when the name carries -select-variables
group	_group_code_from_variable on the data, kept only when the name ends with the matching -{code}
sumlevel	GeoIDFQ.parse(...).sumlevel (geoidfq primary), kept only when the name carries the hierarchy token
scope	the SCOPES key remaining after the survey/year/sumlevel/group/variable markers are stripped

The name's presence/absence of the sumlevel token and group segment is the source of truth for distinguishing the three modes (group-only / variables-only / both) — the long data alone cannot, since all three can produce identical rows. Anchored parsing recovers the other parts from the data first, so the survey's embedded dashes don't make the name parse ambiguous.

Behavior change

load() no longer raises on resources lacking a _morpc block. It raises ValueError only when the name doesn't match the data's survey/year or the residual token isn't a known scope.

Tests

• Replaced test_resource_without_morpc_raises with test_resource_has_no_morpc_block (asserts save() no longer emits _morpc and the resource still loads).
• Added recovery tests for all three modes: test_recover_group_only_no_sumlevel, test_recover_variables_only, test_recover_group_and_variables.
• Full suite: 334 passed.

🤖 Generated with Claude Code