Task: Move classification into refdata catalogs

Make the database the ground truth for classification. The ores.marketdata asset_class and series_subclass C++ enums go away, consumers carry the code string, and both vocabularies live in refdata catalogs that the UI reads for labels.

• refdata.asset_class_code holds the seven-class taxonomy from doc/knowledge/domain/asset_classes.org; instrument_code rows are reclassified off cross_asset and the populate script matches.
• refdata.series_subclass_code ships end to end (table, SQL, domain, repository, service, protocol, registrar, history mapper, Qt client) and is populated.
• No C++ enum in ores.marketdata mirrors either catalog; consumers carry the code string.
• Qt reads asset class and subclass labels from the catalog tables, not from compiled-in names.
• Full local build clean and the full ctest suite green.
• Codegen drift is zero for ores.refdata and ores.marketdata: regeneration reproduces the checked-in output byte for byte.

The catalog becomes the authority first; the code follows it.

1. Refdata. Rewrite asset_class_code to the seven-class taxonomy in doc/knowledge/domain/asset_classes.org: rates becomes interest_rates and cross_asset goes. Reclassify the fourteen cross_asset instrument codes onto their real classes. Update the create, drop and populate SQL, and the RLS policy script that lists the new table.
2. Refdata. Add series_subclass_code as a new entity org and generate its full stack: domain type with JSON and table IO, generator, table, repository, entity, mapper, service, protocol, handler, registrar, history field mapper, and the eventing integration test.
3. Marketdata. Delete the asset_class and series_subclass C++ enums. The domain types carry the code string. Update the two entity orgs, the two generators, the two mappers, the two history field mappers, the domain tests, the feed ingest loop and the import service.
4. Qt. Read the asset class and subclass labels from the catalog tables through the lookup fetcher instead of compiled-in names. Add the series subclass client model, controller, detail dialog and MDI window; register the new lookup type in the refdata and mktdata windows that display it.
5. Ripples. The synthetic IR curve feed, the iam tenant provisioner's asset class literal, the DQ populate scripts and the mixin custom-type list in projects/modeling/cpp_custom_types.org.
6. Verification. Codegen drift for ores.refdata and ores.marketdata must stay at zero, the full build must be clean, and the full ctest suite must be green.

The known follow-up, deliberately left out: the feed ingest loop still special cases FX. That defect gets its own task; this task only stops it depending on the deleted enum.

Manual QA scenarios (scaffolded via compass add test_scenario, run through the QA Validation Runner panel) that verify this task. Link new ones here as they're created; the scenario doc itself links back via its "Verifies task" field.

Local review of the change set, before the PR was raised. Seven findings.

DONE. The classification vocabularies are data, and no C++ enum mirrors them.

What shipped:

• refdata.asset_class_code holds the seven-class taxonomy. rates became interest_rates and cross_asset is gone. The fourteen cross_asset instrument codes moved onto their real classes. The create, drop and populate SQL changed, and the RLS policy script lists the new table.
• refdata.series_subclass_code ships end to end. The set is the modeling org, the generated domain type with JSON and table IO, the generator, the table, the repository, the entity, the mapper, the service, the protocol, the handler, the registrar, the history field mapper, the Qt client model, the controller, the detail dialog, the MDI window, and an eventing integration test.
• The ores.marketdata C++ enums asset_class and series_subclass are deleted. The domain types carry the code string.
• The refdata dialogs read labels from the catalog tables through the lookup fetcher. The market series grid does not: ClientMarketSeriesModel::data returns the raw asset_class and series_subclass codes, so the grid shows the code and not the catalog name. Corrected on review; this bullet claimed the grid read labels, and ores.qt.mktdata carries no lookup fetcher at all.
• Market data validates its classification columns against the catalogs through trigger functions that raise SQLSTATE 23503. market_series validates both columns, feed_bindings validates asset_class, and refdata.instrument_code validates asset_class where it is set.

Acceptance:

• Seven-class taxonomy, reclassified instrument codes and a matching populate script. Met. The populate script holds exactly the seven codes, and neither rates nor cross_asset.
• series_subclass_code end to end and populated. Met.
• No C++ enum mirrors either catalog. Met.
• Qt reads labels from the tables. Met.
• Full build clean and the full suite green. Met. The build finished with exit code 0, and the test run reported 100% tests passed, 0 tests failed out of 73 in 1082 seconds. ores.refdata.core.tests passed in 262 seconds, and every ores.marketdata suite passed.
• Codegen drift zero for both components. Met. check_component_drift.py --all is clean, and regenerate_cmake_component_files.py --all --check reports every component_files.cmake up to date.

Known debt, recorded and not fixed here:

• The ORE vocabulary table in import_service.cpp lost its CORRELATION row, so an import that carries a correlation series now throws std::invalid_argument. The corpus does carry such series. This is the most serious review finding. It needs the market data data-model decision and not a retarget, because a pairwise correlation belongs to no single asset class.
• The DQ publish path and the feed ingest loop still special case their series types. Both leave when the mapping becomes data, which is the work of the "make the remaining hard-coded classification sites into data" task.