Task: Review the rates family's data model

Produce the evidence the parent story needs: walk the checklists over the rates instrument family and record what holds, what does not, and which findings are recorded decisions. The review is read-only. It changes no model, no generated artefact and no mapper.

• Walk every Regime 2 decision check over the rates entity models, with the evidence for each result. Done: the What holds and Findings tables of the investigation.
• Walk the Regime 3 checks over the family's carriers and protocols. Done: `instrument_batch`, `swap_instrument_data`, `trade_instrument`, `rates_fact_variant` and the protocols.
• Spot-check the invariant checks once and state what was read. Done: 25 domain headers, the codecs, the repository entities and the batch.
• Map each finding to a criterion, with file-and-line evidence. Done: thirteen findings.
• Separate a defect from a recorded decision. Done: the soft references, the doubled vocabularies and the flat scaffolding are recorded as component-wide.
• State what was not checked. Done: the investigation states the tree and that no build, test or database run was made.

1. Scope the family: 25 entity models, their 25 tables and the generated C++.
2. Classify the family against the three regimes.
3. Walk the Regime 2 decision checks per model, and the invariant checks once.
4. Read the ORE import and export seam in both directions, because the One-way numeric conversion check is decided there and nowhere else.
5. Walk the Regime 3 checks over the carriers, the batch and the protocols.
6. Re-check every critical claim in the main thread before recording it.

Measured at 3a49daea81, the merge base with main, on a clean checkout. The same table is in the investigation, which is where the findings that use it live.

• The review's shape is the opposite of the FX one. FX had a clean structure and an unmigrated boundary; rates has the best structure of the three families reviewed and a boundary that disagrees with it about what a product contains.
• Thirteen claims changed under the two re-checks and one finding was withdrawn, all recorded in the investigation's Corrections made while checking. Five were found in the main thread — the conversion seam passes at the helper and fails at six sites beside it, the empty cap/floor fact table is a dropped field rather than a product with no fields, the variant is not a Regime 3 violation, the recorded decision that the variant was removed is false, and a floor is corrupted in the reverse direction. The other eight an independent verifier found by opening every citation: the soft-reference count, the leg-direction finding's two citations, that finding's claim that the import always supplies a direction, three nats_call citations, the task's copy of the money count, the batch's array count, a sequence-line citation, and a scored regime that contradicted its own checklist row. That is the step which should not have needed a second agent.
• A floor becoming a cap was found on a second pass, after the strike finding was written. Following a lost column in both directions is what raised it.

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

The review is recorded as Investigation: Rates family data model review.

Regime 2 scores 2 of 5 with debt High, and Regime 3 scores 5 of 5 with debt Low. The split is the finding: the family's structure is the best of the three reviewed — identity is the key, the legs and their economics are typed rows, the variation is generated per-type tables, the money is exact at rest — and its boundary is where the debt sits. Thirteen findings, three of them Critical: a short forward rate agreement stores as a long, an inflation swap cannot be stored at all, and a floor returns as a cap.