Input-side nesting-projection classification (NestingType mirror)

A directiveless SDL input type nested under a table-bound parent (a @table input, or a jOOQ-record @service param via the R336 flatten, now Done) is semantically a projection of columns on the parent’s table: its fields resolve against the parent’s TableRef, it has no table of its own and no Java backing. Today it classifies as GraphitronType.PojoInputType with fqClassName = null (the bindings.resolveInput(name)-empty branch in TypeBuilder.buildNonTableInputType). Calling a column-grouping projection a "POJO" is a misnomer that leaks a reflection fallback into the model and the LSP (hover + inlay).

R337’s plan was a type-level relabel: introduce a new variant and assign it instead of PojoInput(null). The relabel is codegen-neutral, the nested grouping’s record class still emits via HasInputRecordShape, and both flatten paths (the @table-input classifyInputField recursion and the R336 InputBeanResolver recursion) read the raw graphql-java Map, never the typed record, so the work really is just "add a variant + the two exhaustive GraphitronType switch arms (projectTypeClassification, projectType) + an LSP label/hover + tests."

But R97 (via the R327 fold) settles that input classification is contextual, a function of the consuming field/coordinate, not a global property of the type, and names the null-backed PojoInputType explicitly as the contextless artifact that "exists only because buildInputType is asked to classify globally with no consumer in view. Per-coordinate, it cannot occur." R97’s fix is to dissolve the per-type verdict (consumer-derived resolution, grounded in R333’s coordinate-lowering), not to add another per-type label. R337’s mechanism is the exact type-level altitude R97 diagnoses as wrong, and R97 records a withdrawn attempt that failed because "the type-level demotion throws away context the coordinate already has." Adding R337’s variant now is model surface (a GraphitronType permit, a TypeClassification permit, HasInputRecordShape / InputType membership, LSP labels, pinned tests) that R97 / R333 would then have to remove: negative work against the agreed direction.

R97 is scoped to query-filter binding, @table-on-input retirement, and the mutation write-target axis; it does not explicitly name the honest surfacing of a nested-grouping projection (the LSP hover / inlay and the TypeClassification label that today reads PojoInput). If R97 / R333 land their per-coordinate model but leave that surfacing on the null-PojoInput label, revive R337 narrowly as "surface the projection honestly on the lowered coordinate," with no new per-type variant. Until then this file is a redirect.

Backlog tombstone (per workflow.adoc): kept as a redirect during R97’s lifecycle, depends-on R97. Delete this file when R97 reaches Done if the surfacing residual is subsumed there; otherwise re-scope to the residual only at that point. Out of scope regardless: the functional flatten (R336, Done), how nested fields resolve to columns, and the output-side NestingType.