70. Keep Repetition Records Across Unchanged Document Parses

• A root memo hit on an unchanged Document.Parse previously published an empty run map, losing the long repetition's next-edit resumption. Unchanged parses now keep the map from the same edit generation; rebuilding a repetition can replace its record in that map. Only an edit advances the generation. A parse after an edit still creates a new map; two unparsed edits, aborted execution and edit-log rollover retain the existing invalidation rules. The unchanged path copies no records and allocates no replacement map. All engine backends apply this; generated parsers have no Document API.

• Five alternating pairs, Apple M3 Max, Go 1.27.1, 200 ms per sample, compare baseline 95f4e3a with this same-generation run-map retention. The focused workload has 100,000 ab\n lines and replaces one character in the middle. Compilation, document construction, initial parsing and the optional unchanged parse are outside the timer; Edit plus the final Parse are measured. With the redundant parse, 0 → 99,999 elements resume on all engines and both units. Code-point medians are:

  Backend	Time, before → after	B/op, before → after	Allocs/op
  Closure	6.90 → 0.778 ms (−88.7%)	48.83 → 1.15 MB	40 → 11
  Recursive VM	8.03 → 1.413 ms (−82.4%)	57.76 → 10.08 MB	73 → 44
  Iterative VM	9.06 → 1.417 ms (−84.4%)	57.76 → 10.08 MB	81 → 52

  Byte-position timings fall 89.1%, 83.1% and 84.9%, respectively, with the same allocation counts and almost identical bytes. Without the redundant parse, all six controls retain 99,999 resumed elements and unchanged allocation counts/bytes (at most 4 B/op difference); time medians range −4.9% to +5.8% with overlapping sample ranges. This restores later edit reuse; it does not measure the cost of the unchanged parse itself or claim a batch speedup.

• The existing minilang Document/fresh and 50,000-record CSV incremental controls also have five alternating 200 ms pairs. All nine sample ranges overlap; median times change −3.4% to +0.9%, allocations stay unchanged and B/op varies −0.8% to +0.02%. Their edit schedules do not include a redundant Parse, so these are regression controls rather than evidence of the new-path speedup.

• Retention extends the lifetime of valid current-generation records across unchanged calls, as needed for the next edit; it does not archive generations. Tests compare values, spans and errors against fresh parses after repeated no-edit calls, nested edits with shifts in both directions, ordinary failures and recovery, multi-edit gaps and edit-log resets. A primed root-memo trace panic clears the shared run map/memo/scratch, then retry and edit reuse work. Existing action/depth and provisional-left-recursion invalidation tests pass.

• Reproduce with go test ./internal/engine -run '^$' -bench '^BenchmarkDocumentReparse$' -benchtime=200ms -benchmem and go test ./bench -run '^$' -bench '^BenchmarkIncremental(Long)?$' -benchtime=200ms -benchmem. Build separate baseline/candidate binaries with the same benchmark source; restore only internal/engine/document.go from the baseline via a Go overlay and alternate their runs. Raw samples remain local. The next full benchmark checkpoint follows the core compiler analysis optimization group.