Task: Give a market series a set of asset classes

Replace the single not-null market_series.asset_class column with a junction table, so a pairwise correlation can reference the two asset classes it relates. Update every consumer of the column, and confirm the refdata catalogue descriptions read correctly.

• ores_marketdata_market_series_asset_classes_tbl exists and is tenant-scoped, and a series carries any number of asset classes.
• market_series.asset_class is gone: the column, its check row, its generator, and every reader of it.
• An import writes one junction row per class the classifier returns, so a cross-class correlation writes two.
• The Qt market series grid no longer shows a raw code the series does not carry, and the column and its filter come back sourced from the junction under the sibling task. Amended on review; this bullet claimed the grid would show the catalogue name, which the change does not deliver because no read path for the junction exists.
• Every name and description in the refdata catalogues this work touched reads correctly.

Unit 4 of the story plan, then Unit 5. Unit 4 runs before the sibling task's Unit 1, so that no step in the sequence needs a guarded front() or an invented filler for the empty class set.

Unit 4, the junction. A new ores.codegen.junction org, ores.marketdata.market_series_asset_class_junction, modelled on ores.iam.account_party_junction for the uuid-keyed shape. Left is market_series_id, right is asset_class_code, both soft FKs. The profile binds under * Flags, not in front matter, because load_org_junction_model reads front matter through a fixed allowlist that does not include profile. Generate it, then drop asset_class from market_series.org and regenerate.

Consumers to update: import_service.cpp, feed_ingest_loop.cpp, the DQ publish SQL, ClientMarketSeriesModel, MarketSeriesMdiWindow, MarketSeriesPickerDialog, and the domain tests.

Unit 5, the catalogues. The populate SQL for the catalogues this work touched already carries name and description for every row. This unit reads them, confirms each reads correctly, and fixes any that do not.

Three deviations from the plan, all deliberate, all recorded here rather than absorbed.

The plan expected the domain type to carry a collection of asset class codes, and the Qt grid and its filter to read that collection. Neither shipped. A series reads its classes from the junction, and nothing else holds them, so a client cannot know a series' classes without a second request — and no protocol serves one. Junction messaging is hand-written, not generated, so serving one means writing a protocol, a handler and a client fetch by hand.

The three items this defers, all to the sibling task: the grid's Asset Class column, the toolbar filter combo box that filtered it client-side, and the New Series form's asset class combo box. The column and the filter are a regression, not a gap: both worked before this change, showing and filtering on the raw code.

The alternative was to keep the column reading a value the series no longer has, which cannot compile, or to leave a filter combo that filters on nothing, which is a control that lies. Deleting them and restoring them from the junction is the smaller honest step.

The plan's third expectation holds as written: import_service writes one junction row per code the classifier returns.

The consumer list missed a populate script, and the recreate caught it. marketdata_market_series_fomc_populate.sql seeds the two FOMC segment series and named asset_class in its column list. The column is gone, so the insert failed. setup_schema.sql sets ON_ERROR_STOP, so the populate phase aborted there, and compass db recreate returned 1. The failure hid its own cause: the metadata phase runs after the schema phase, so a non-zero exit left no metadata row and no ores.env_type, and the schema itself looked complete because refdata populates before marketdata.

The fix removes the column and writes one junction row per series. A full db recreate now completes with no errors in 1m 25s, and both the series and junction rows are present.

The consumer list missed a second reader, and the build caught it. RateCurvesMdiWindow.cpp filtered its rows on s.asset_class ! "interest_rates", which no longer compiles. The row does not carry the class, and the junction has no read path, so the filter reads =series_subclass = "yield"= instead – the catalogue's rates-only, curve-shaped subclass. That is narrower than the old test, which also admitted the spread, FRA, xccy and volatility subclasses.

That filter was also hiding the window's own content. Its second clause called has_point_dimension, and the registry has no entry for RATES or YieldCurve. Those are the two series types the synthetic feed and the republish service write, so the window exists to show them, and both were filtered out. has_point_dimension answers false for a type it does not know. decompose_key treats an unknown type the same way, and folds every segment into the qualifier with no point. The new filter drops the call, so the seeded curves appear. No other code in the tree calls has_point_dimension.

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.

Unit 4 shipped. market_series.asset_class is gone, and a series carries its classes through ores_marketdata_market_series_asset_classes_tbl, generated from a new ores.codegen.junction org.

What changed. The junction org models market_series_id against asset_class_code, both soft FKs, under the tenant-scoped-junction profile. Dropping the column took its check row, its generator and its insert-trigger validation with it. import_service writes one junction row per code the classifier returns, so a cross-class correlation writes two. feed_ingest_loop, the DQ publish SQL, the domain tests and six Qt files followed. Two stale component_files.cmake were regenerated.

What the change surfaced. Three defects, all recorded in * Notes above. The consumer list missed a populate script, and the db recreate caught it. It also missed a second reader, and the build caught it. That reader, RateCurvesMdiWindow, had a further latent fault: its has_point_dimension clause dropped both seeded series, because the registry holds no entry for RATES or YieldCurve and the predicate answers false for a type it does not know. The grid was empty before this work began.

Verification. compass db recreate -y -k completes clean in 1m 25s, with both the series rows and the junction rows present. The full build reaches 100% with no errors and no warnings. The suite passes 73 of 73 in 1126s.

The corpus classifier tests ran. ores.marketdata.core.tests passed 697 assertions across 191 test cases, and all four corpus cases appear in the run. every_distinct_series_key_in_the_ore_corpus_classifies clears its 90000-key floor, so the walk found the corpus rather than an empty tree. The suite randomises case order, so the walk's cost lands on whichever case calls the survey first, about 18s; the rest reuse it.

Not covered. No automated test reaches the Qt windows. The RateCurvesMdiWindow filter change is verified by compile, link and inspection only.

Deferred to the sibling task. The grid's Asset Class column, its toolbar filter and the New Series form's class combo box. Nothing serves the junction over the wire, so restoring them means a hand-written protocol, handler and client fetch.