Classify the imports the target-compilation kernels' hashes do not yet cover #1497

#1493 (for #1492) adds a check that walks every module each target-compilation kernel passes to source_hash, the kernel's own module included. It requires every first-party module those modules import to be hashed by that kernel, or recorded with a reason. source_hash hashes only the files it is given and does not follow imports, so an import nobody hashes is code a kernel can run without its node key changing.

The kernels checked, in groups that share one hash:

• UK:
  • the full graph's uk.full.target_compilation@1, target_selection@1 and measures@1;
  • uk.full.problem@1;
  • the national graph's national_targets@1 and national_problem@1.
• Transport:
  • targets.compile@1, takeup.compare@1, takeup.gap@1, transport.quantile_map@1 and @2, each alone;
  • transport.create@1 with @2.

On the first run the check found imports that no review has classified yet. #1493 records them as they are, in an UNCLASSIFIED_IMPORTS mapping. Each importer -> imported edge names the groups that leave it unhashed:

• UK: 66 modules, 112 edges, in packages/microcosm-build/tests/engine_free/uk/test_uk_target_compile_modules.py.
• Transport: 35 modules, 74 edges, in packages/microcosm-build/tests/engine_free/shared/test_transport_compile_modules.py.

The check fails on an unhashed edge a group meets that isn't recorded for that group, and on an edge recorded for a group that no longer meets it. So the mapping can only shrink, a new import of unhashed code fails at once, and removing a module from one group's hash fails even where another group records the same edge.

Classify these first: code the kernels run on every build today.

• uk_runtime.national_calibration, which uk.full.measures@1 runs through full_measure: CalibrationFrameAdapter, inject_measure_inputs and drop_injected_measure_inputs. The national kernels already hash this module; the four full-graph kernels don't.
• uk_runtime.rowwise_dataset.ladder_clone_index_column, which uk.full.measures@1 runs through full_measure. No UK kernel hashes rowwise_dataset.
• uk_runtime.graph_population.context_frame, which rebuilds the population in the full measures and problem kernels and in the national problem. None of the six UK kernels hashes graph_population.

Goal: classify every entry, then delete the mappings. For each entry, one of:

• Hash it in the kernels that run it. Over-keying is safe, but each module hashed by the UK target kernels also re-runs the UK solve when it is edited. That works against reusing the solve between two builds (UK local-area release: publish a K=15, 150,000-household dense dataset in at most two builds #1301).
• Move what the kernel runs into a small module that the kernels hash. Include geography_constants in the target-compilation cache keys #1493 does this for calibration_run._ledger_provenance and _validate_band_edge_registry, now in uk_runtime/ledger_surface.py.
• Record why the kernel doesn't need it hashed, for example that the import serves another kernel or the graph builder in the same file. One recorded entry already carries such a note: graph_national -> calibrate.solve, which only uk.full.national-gates@1 runs.

What the other entries are. Most fall into three groups:

• imports of a kernel's own module that serve the graph builder or other kernels in the same file, such as graph_national -> graph_build, graph_calibration, staging_v2 and logbook_adoption;
• the UK measurement code's other imports, such as measure_simulation -> frs_spine, frame.adapters.policyengine_uk, lcfs_consumption and uc_target_measurements, many of which probably run when the measures kernel runs;
• the microcosm.frame and microcosm.graph modules, and the imports of country_spec's loader (spec_engine.*).

Out of scope: the modules #1493 hashes: calibrate.target_selection, calibrate.matrix, calibrate.artifacts, calibrate.solve (in the full problem only), uk_runtime.weighted_integrity and the new uk_runtime.ledger_surface.