@oneOf definition augment defeated by descriptions quoting the definition; federated SDL fails Apollo composition

OneOfDirectiveSdl.augment (generators/schema/OneOfDirectiveSdl.java) reinstates directive @oneOf on INPUT_OBJECT into the federation SDL after ServiceSDLPrinter.generateServiceSDLV2 strips spec-built-in definitions. Its already-present guard is a substring check, sdl.contains("directive @oneOf"). The federated fixture’s Query.filmsByOneOf field description (authored in graphitron-sakila-example/src/main/resources/graphql/federated-schema.graphqls) quotes the literal string directive @oneOf on INPUT_OBJECT while explaining the R283 mechanism, and that description prints into the emitted SDL. The substring check matches the description text, the append is skipped, and the emitted file (target/generated-resources/graphitron/no/sikt/graphitron/generated/federated/schema.graphqls) carries the @oneOf application on FilmOneOfFilter with no definition. The comment explaining the fix defeats the fix.

Both arms are affected identically: the codegen-side file arm (SchemaSdlEmitter.printFederationServiceSdl calling OneOfDirectiveSdl.augment) and the generated runtime arm (the <outputPackage>.util.OneOfDirectiveSdl mirror emitted by OneOfDirectiveSdlGenerator, whose logic is single-sourced from the same class), so the runtime _Service.sdl a router introspects is equally broken for any schema whose descriptions quote the definition. The shipped fixture hits it; a consumer schema documenting its own @oneOf usage can hit it the same way.

The federation round-trip tests parse the emitted SDL with graphql-java, which knows @oneOf intrinsically whether or not the definition is present. Only a real Apollo composition run exposes the gap; none exists in the build today (that check is R298’s work package 1, and this bug is its first scalp, found during R298 research rather than by CI).

Replace the substring guard with a check that only matches an actual definition, e.g. a line-anchored match ((?m)^directive @oneOf\b) or, better, detect the definition on the parsed document rather than the printed string. Keep the file arm and the generated runtime arm single-sourced as today. Add a guard test whose schema has a description quoting the definition text (the shipped federated fixture already is one; asserting on its emitted SDL that a line-anchored directive @oneOf definition exists would have caught this). When R298’s compose check lands, it covers this class of failure wholesale.