Task: Generate the oresmd grammar from the models

The oresmd grammar that ships is the one the models declare. A URI names every coordinate dimension with its own query key, writes every key the grammar defines for its type, and names no default by silence. point is gone from the identifier, the parser, the writer and every declared test case.

• No query key is silently defaulted: the parser materialises a grammar default and the writer emits it, so model is always present for type=vol and metric and quote always for type=quote.
• No coordinate is a comma-joined string: point is gone from the identifier, the parser and the writer.
• ATM is a delta and not a strike.
• Every URI in every model's test tables uses the new spelling, the generated tests pass, and a URI carrying point is refused.
• doc/knowledge/domain/market_data_urn.org describes the grammar that ships.

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.

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 oresmd grammar that ships is the one the models declare. point is gone from the identifier, the parser, the writer and every declared test case, and every coordinate dimension is a query key of its own.

Delivered:

• The identifier drops point; the surface keeps volatility_surface_point as its typed container rather than gaining six top-level members. The plan asked for the struct's deletion too; the acceptance asks only that no coordinate be a comma-joined string, which this meets, and the deletion is left as optional.
• The parser refuses point with a message naming the replacement keys, reads each declared coordinate by its query-key name, and refuses a coordinate the resolved quote type does not carry. That check is generated from the model's coordinates column, so the grammar is enforced by construction.
• The parser materialises the quote-type default for a type=quote URI that names none (spot for FX, equity and commodity; cds for credit), and swaption or option for a type=vol URI that names none.
• The writer emits every set key, writes model even when it holds the default, and skips an empty ccy for a quote type the model marks no_ccy.
• The projections' key builders read the typed members. split_point, the derived point and compose_point are deleted.
• to_series_uri() clears the declared coordinates and keeps the surface, so model is identity. The IR std::visit with three special cases is replaced by members the model declares. This is the fix for the corpus's lognormal and normal vol colliding: the coverage walk now names and round-trips all 485,606 of its keys, where main round-tripped 481,311.
• ATM is a delta: a surface's value slot is strike for a number and delta for a moneyness, and smile is its own dimension.
• Every point-bearing URI in the twelve models' test tables and in the templates moves to the named keys, and the codegen gate is rewritten to check the new spelling. It now fails a positive test case that still spells point, a coordinate the resolved quote type does not carry, or an undeclared key.
• doc/knowledge/domain/market_data_urn.org describes the grammar that ships.

What is verified, and where:

• The oresmd suite: 1,225 assertions in 363 cases, all green, at c8aeeabbeb.
• The marketdata core suite and all 89 ctest suites: green.
• The codegen suite: 1,050 passed.
• The whole-corpus coverage walk: 485,606 keys named and 485,606 round-tripped.
• compass services restart: 24 of 24 services running, none failed.

What is not met, recorded rather than glossed:

• metric is not materialised. The acceptance asks the parser to write it always for type=quote; it is still supplied at projection time from the quote type's own metric. Materialising it needs a per-quote-type metric constant the models do not carry for every class – credit's BASE_CORRELATION has no metric enum member – so it is left to a unit that can extend the enum.
• The parser does not require a coordinate of a type. A type=vol&quote=capfloor URI with no strike parses; the projection then refuses to emit a key. The old grammar's point-arity errors are not re-stated as a parser-side rule, because no model declares which coordinates a type requires and inventing one would put a convention back in the parser that the declaration is supposed to replace.
• The twelve models still carry a point_keys column, the ordered keys the old point value corresponded to. Nothing reads it now: the parser is hand-written per family and the codegen gate checks the named keys instead. Removing the column and the loader's validation of it is a mechanical subtraction left for a follow-up, recorded rather than done because it changes nine tables whose separator rows a mechanical strip mangles.
• The whole-corpus import fails on market_20160205.txt. The series identity now drops credit's tenor – which the model declares a coordinate – while the observation's point_id is still the registry's ORE-form split, which is empty for the four-segment INDEX_CDS_OPTION/RATE_LNVOL/<INDEX>/<TENOR> term vol, so two term vols of one index collide on an empty coordinate. Moving the stored coordinate to the declared keys is next unit's work (task 4204259C).