[OmarAlJarrah]

Summary

Four facts a source states about a type node had no field in the IR. In two of them the IR said two things at once.

• Open unions (ir: a union can't be marked open #664). Union gains Closed, as Enum has it. A JSON Schema oneOf or anyOf admits only what matches a listed branch, so every union the OpenAPI compiler lowers is closed. A Smithy union, a protobuf oneof and a GraphQL output union are open: a consumer must represent a variant it does not know. docs/ir/unions.md#open-unions states the default for each format, and the planned lowerings follow it. The capability matrix gains an open-unions row. The OpenAPI compiler gains a promotion target, union.open, read by presence as enum.open is. It has no default key, because no spelling has become common. testdata/conformance/openapi/matrix.md therefore records why no corpus spec can witness the row, and names the test that holds the promotion instead.
• Input-only unions (spike: evaluate first-class IR representation for input-only unions and non-finite numeric values #25, case 1). Union gains InputOnly, as Model has it. A GraphQL @oneOf input object is the source: an emitter projecting separate input and output types reads the flag instead of a format-named marker.
• Non-finite values (spike: evaluate first-class IR representation for input-only unions and non-finite numeric values #25, case 2). ValueKind gains inf, neginf and nan, each with no payload. A protobuf [default = inf] was a construct the IR could not hold, and every target with an IEEE float has a constant for each value. They are kinds rather than a flag on number, so a consumer switching on Kind meets them as kinds to handle, never as a number with an empty Num. BigVal stays a finite JSON number, and irverify's payload rules hold the new kinds with no rule of their own.
• Literal value types (ir: a Literal carries no value type or format #668). Literal gains ValueType, read as enumValueType reads an enum's: the type and format the value fits, else the type, else the value's own kind. {type: integer, const: 80.0} is integer and {type: number, const: 80} is number; they were one Value. A date const is date. A type or format the value fits no longer draws an info report or an Unmodeled entry. One it does not fit still draws both, and so does a type set naming several types. irverify holds a ValueType to the declared kinds (ir/unknown-prim-kind at the field).
• A closed rest beside patterns (ir: a closed object with patternProperties also carries a catch-all typed as the top type #987). AdditionalProps.Value becomes *TypeRef. patternProperties beside additionalProperties: false used to type every key no pattern matches as the top type, while Additional said no such key was allowed. That catch-all now holds the patterns and no value. docs/ir/models.md#catch-alls states the three shapes. irverify reports the two that would say two things: a catch-all typing no key (ir/catch-all-empty), and a closed model's catch-all with a value (ir/catch-all-closed).

On the golden spec (spec @319dc756), against this PR's base, the only change is the two new members on its 974 unions. The diagnostics are identical. The spec writes no const and no patternProperties: its 195 literals are null or list members of an enum lowered as a union. That is the rule-3 ranking these issues carry.

Closes #664
Closes #668
Closes #987
Closes #25

Breaking

IRVersion moves to 0.31.0, the generation after main's 0.30.0 (#1327). The branch's earlier commits name 0.28.0 and 0.29.0, from before #1325, #1333 and #1327 landed; its merges of main renumbered it. A consumer pinned to 0.30.0 refuses the document for its unknown members, closed and inputOnly on every union and valueType on a literal. One that decodes AdditionalProps.value as required fails on a closed pattern object, where the member is absent. In Go, AdditionalProps.Value is a *TypeRef. An open PR claiming the same version moves its own constant on if it merges after this one.

Stack

Base: #1355, on #1354, #1351 and the emitter-design stack. main (with #1229) is merged down the whole stack, so this diff is this PR's alone.

Test plan

• go test on the touched packages: ir, ir/irverify, ir/irfacts, compilers/openapi/internal/{schema,lowering,merge}, compilers/openapi/e2e, internal/conformance, cmd/morphic, cmd/morphic-census. The goldens, the shape pin and the census were regenerated with -update.
• New tests: schema.TestLiteral_ValueTypeReadsTheTypeAndFormatItsValueFits (13 rows: fit, misfit, untyped, several types, composite, null); schema.TestObject_CatchAllBesideAnotherFactStaysModel gains the two pattern rows; irverify.TestCheckCatchAlls_ReportsACatchAllSayingTwoThings; irverify.TestVerify_ALiteralsUndeclaredValueTypeIsAViolation; lowering.TestPromoterUnionOpenness_OpensOnAMappedKey. The maps conformance case gains ClosedPatternMap.
• Probes. The issue's three schemas compiled: integer, number, date, with no diagnostic. A misfit type: integer, const: x keeps type and reports it, and type: object beside an object const keeps nothing. The golden spec was compiled at the base and at this head, and the type and diagnostic differential is as stated above.
• Push checklist: subject within 72 characters with the number, no artifact hits, no conflict with main or the base, skills_lint clean.
• After main gained fix!: type the local time formats and keep time's offset #1325, the whole stack was merged forward and every touched package rerun: ir/..., compilers/..., cmd/..., engine/..., pass/..., internal/....
• CI.