72. Reuse Structurally Equal Duplicate Capture Types

• Every duplicate capture label previously called the general type join, which formats, sorts and deduplicates nested record fields. Repeated references to the same recursively inferred record allocate these strings again for each occurrence in every inference round. The capture merger now compares immutable type structures and reuses the previous equal basic, named, list or record type. Ordered nested joins and named-node kinds are compared conservatively; top-level optionals and joins still use the general normalizer. Capture availability remains the logical OR of both occurrences. This changes shared type checking before compilation or generation, not the parsing runtimes or generated matching code.

• Five alternating parent/candidate pairs with a 200 ms benchtime compare baseline 0ab91da with this change on Apple M3 Max, Go 1.27.1. Grammar parsing and Program setup are outside the timer; each iteration runs the complete checkTypes, including dependency analysis, recursive inference and diagnostic checking. The recursive fixture repeats "😀" x:main in main; the record fixture repeats x:item, where item captures a Match and a list of nested records.

  Shape	Occurrences	Time before → after	B/op before → after	Allocs/op before → after
  Recursive	1,024	423.966 → 61.126 ms (−85.6%)	1.188 GB → 32.16 MB	8,204,256 → 274,551
  Recursive	8,192	4,096.058 → 511.795 ms (−87.5%)	9.528 GB → 274.67 MB	65,628,207 → 2,139,000
  Record	1,024	1.921 → 0.456 ms (−76.3%)	2.065 → 1.067 MB	51,277 → 4,219
  Record	8,192	16.432 → 4.398 ms (−73.2%)	16.500 → 8.505 MB	409,698 → 32,912
  Record	32,768	65.064 → 17.327 ms (−73.4%)	68.765 → 36.784 MB	1,638,524 → 131,239

  All five time ranges are disjoint. Bytes are cumulative allocations over one complete checker invocation, not retained heap. Current bounded recursive inference scales approximately linearly in capture count in these samples; an older quadratic description is not a current complexity claim. Structural comparison still traverses nested types and expression checking still processes every occurrence in each inference round.

• Separate single-iteration diagnostics at 32,768/131,072 recursive captures take 14.229 → 2.025 s and 55.894 → 8.210 s, respectively. At 131,072, cumulative allocations fall 152.647 → 4.578 GB and allocation counts 1,050,020,237 → 34,088,466. These reproduce the original large case and are diagnostic observations, not five-sample estimates. The remaining inference/AST traversal work is still substantial.

• Mixed A/optional B duplicates retain general join normalization: 1,024/8,192/32,768 pairs change +0.1%/+0.3%/−0.6% in time, with overlapping ranges and essentially unchanged allocation counts/bytes. Distinct-label record controls change −0.3%/−0.2%/−0.8%, also with overlapping ranges. Unchanged calculator/minilang/JSON grammar AST controls change +0.6%/approximately 0%/−4.2%, all with overlapping ranges and identical bytes/allocation counts. These controls do not establish runtime speedups or promise savings for growing heterogeneous joins.

• Type regressions retain exact public types, recursive widening, nested records/lists, optional and choice availability, repetition-local captures, nil actions and capture-reading actions/predicates. Duplicate-label error positions remain exact across Unicode sequences at several sizes. A pairwise oracle compares the capture merger with the general join for all availability flags, including reordered joins, nested optionals, never/any/nil and equal names with differing node kinds. Independent source review checks normalization, ordering and immutable type reuse.

• Reproduce with go test ./internal/engine -run '^$' -bench '^BenchmarkCaptureChecking$' -benchtime=200ms -benchmem. Build parent/candidate binaries with the same benchmark source, restoring check.go and types.go from 0ab91da through a Go overlay for the parent. Run five alternating pairs for recursive sizes 1,024/8,192 and all record/mixed/unique/control cases, without concurrent tests/builds/probes. Large recursive cases use separate single-iteration diagnostics. Raw results stay local; the full suite checkpoint follows this core compiler analysis optimization group.