Task: Give the FX accumulator range bounds their own table

Replace fx_accumulator_instrument.range_bounds_json with a child table. The column holds the ORE RangeBounds list as JSON text, so a child record has no table and the column carries an encoded document. The Regime 2 check asks for rows, and the Regime 3 check forbids an encoded string field.

• A child model holds a range bound: an optional start date and the optional levels RangeFrom, RangeTo, Leverage, Strike and StrikeAdjustment. Each level is exact decimal where it is money.
• The child table is keyed by the trade and ordered, so the list order survives the round trip.
• A mapper test holds the digits of a range-bound strike through the round trip.
• An accumulator with no bounds is refused at the boundary with a stated error, not silently exported: the RangeBounds wrapper is optional in the schema, but the binding writes it unconditionally and the schema requires an entry inside it, so a trade with no bound has no valid document.
• range_bounds_json is gone from the model, the SQL, the domain type, the entity, the protocol and the mapper.

The unit is two changes, because the FX family has nowhere to carry a child today and the second change cannot be made without the first.

1. The carrier. Every other family returns a struct that holds the instrument beside its children: commodity_instrument_data, swap_instrument_data, bond_instrument_data and so on. The FX family is the exception: map_fx_instrument returns fx_instrument_variant, one aggregate that carries an entity, which The atomised entity forbids. The carrier becomes one typed array per product, joined by trade id, and the variant is deleted. That is its own unit: Give the FX family typed arrays, not a variant.
2. The child. Author the child model, remove range_bounds_json from the accumulator model, regenerate, and map the child in both directions into one more typed array on the carrier. Delete the JSON helpers, range_bound_to_json and parse_json_range_bounds. Then recreate the database and run the trading and ORE suites.

The carrier is a mechanical refactor over ten production files and the tests that unwrap the variant, with no behaviour change and no schema change. The child is the schema change.

Every number this task claims about a tree, with the command that produced it and the commit it was measured at. A number without both is not a measurement, because a clean-up record describes a tree and trunk moves under a branch: a consumer count, a gate's component count and a suite's size all change with no code changing, and the reader of a bare number cannot tell which tree it came from. Where the raw output matters, put it in an evidence/ file beside the task and name that file in the Command column.

• The evidence for this task is Investigation: FX family data model review, finding 2. The model states the gap itself: "The list has no child table, so it lives in JSON text."
• The levels pass through float text on the way in and out today, so this unit also removes a precision loss.

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.

A second review checked the fixes on the new head. It confirmed every fixed finding with a call site or a test, and raised four more:

• The refusal escaped the manual export path. ore_commands::process_export called the exporter bare, so one boundless accumulator ended the shell session instead of reporting the failure. Fixed: the command catches the refusal and reports it, as the package workflow already did.
• The twelve-decimal text had not reached the SQL. Fixed: the SQL regenerated and is committed.
• The refusal stated the schema more strongly than it says. The RangeBounds wrapper is optional; what makes the refusal necessary is that this binding writes the wrapper unconditionally and the schema requires an entry inside it. Fixed in the model, the header, the test and the acceptance line.
• No test asserts the digest reads the child, and the repository test called the digest reader the export reader. The second is fixed: the test now asserts both readers. The first stays open: the digest is asserted through the generated shape, and a per-entity digest test belongs to the digest's own unit.

The accumulator's RangeBounds list is a child table. range_bound is keyed by the trade and the entry's ordinal, holds the optional start date, the four money levels as exact decimal and the leverage as a continuous double, and carries the two foreign keys the other instrument children carry. range_bounds_json is gone from the model, the SQL, the domain type, the entity, the protocol, the shell command and both mappers, and the hand-written JSON serializer and parser are deleted.

The four money levels cross the ORE boundary through moneyFloat, so the digits the document states reach the decimal column. Two goldens recorded the old float spelling and now record the document's own: 1.04999995 became 1.05, 1.10000002 became 1.1, and a stated 0.0 stays 0.0.

An accumulator that holds no range bound is refused on export with a stated error, because the ORE schema requires at least one and a document without it cannot be read back.

The child is keyed by the trade alone, so the equity accumulator can move onto the same table in its own unit; it still writes JSON text today.

• The child is trade-keyed, not FX-keyed. A RangeBounds list belongs to the trade its document states, and the equity accumulator states the same list, so one table serves both. Keying it to the family would duplicate the shape.
• Leverage stays a double. It is a multiplier, not a money amount, so the continuous rule applies and the column is a binary double. The four price levels are money and are exact decimals.
• The levels are money leaves in the binding. rangeBound's five levels were plain floats, so the document's digits were lost before the mapper ran. They are moneyFloat now, added to the reapply script with the rest of the family.
• A stated level the binding never carried is null, not zero. Every level is optional in the schema, so an absent level is a null column and the reverse mapper writes no element back.
• The trailing zero is the one spelling the round trip normalises. A stated 0.0 reaches a decimal column as zero, so the golden records 0.0 only where the binding carried the text; a value that passes through the decimal writes its canonical spelling.