perf(strawberry-shake): index type descriptor lookups to remove O(N²) mapping

[cliedeman]

Root cause

TypeDescriptorMapper (StrawberryShake C# generator) scaled super-linearly with operation count: profiled at ~1.9s for 200 ops and ~4.3s for 275 ops on a real 672-op client. Three field resolutions performed linear scans over collections that grow with every operation:

• AddProperties → GetFieldTypeDescriptor looped over every operation and then did typeDescriptors.Values.First(t => t.Model == fieldType), an O(types) reference scan per field. Profiling showed each call scanned ~110–160 operations on average (1.4M–3.4M operation scans in total) — the dominant cost.
• AddInputTypeProperties → GetInputTypeDescriptor did typeDescriptors.Values.First(t => t.Model.Type.Name == …) per input field.

Fix

Build dictionaries once per Map call and resolve in constant time:

• descriptorsByModel maps each output type model to its descriptor (reference keyed), replacing the First() scan in the operation-scanning fallback.
• interfaceDescriptorsBySelectionSet maps an interface output type's selection-set node to its descriptor. A field whose result is an interface is resolved directly by the selection set its syntax node points at, which uniquely identifies the registered interface descriptor — eliminating the per-field operation loop. The original loop is retained as a fallback for any field whose selection set is not in the index, so behavior is unchanged.
• descriptorsByTypeName maps input descriptors by GraphQL type name (first insertion wins, mirroring the prior First() enumeration order).

Selection-set nodes and output type models are compared by reference (the analyzer interns selection sets; the models define no custom equality), so the indexes return the identical descriptor instances the scans returned.

Correctness (generated output byte-identical)

Verified the broad test slice (StarWars*, ResultType*, Operation*, Entity*, DataType*) on this machine produces the exact same pass/fail set and byte-identical generated sources with and without the change (identical SHA over all generated files). An empirical equivalence check over the full 672-op client confirmed the selection-set index returns the same descriptor as the operation-scanning loop for all 33,781 field resolutions (0 misses, 0 mismatches). The 51 pre-existing snapshot failures on this machine are an unrelated BOM/encoding environment issue present on main; the change neither adds nor removes any.

Before / after — TypeDescriptorMapper.Map (672-op ILOPS client)

N (ops)	Before	After	Speedup
200	1933 ms	454 ms	4.3×
275	4307 ms	435 ms	9.9×

The mapper is now flat with operation count instead of super-linear.

Relates to #17

— Claude
🤖 Generated with Claude Code