69. Bound Persistent Variable Binding Histories

• Every binding previously prepended a binding even when the name already existed. A long repetition retained all shadowed values, and reading an outer variable traversed them all. The current environment now holds one binding per name. Replacing a changed non-head value copies its prefix and shares the suffix, preserving saved environments for rule scopes and backtracking. Binding an equal scalar value reuses the environment. A monotonic name hint skips searches for first bindings and records names from the first binding, including abandoned branches. Closure, both VMs, generated Go/typed Go and TypeScript use this logic.
• For a stream with three bindings per record and three distinct names, the old current list has 60,001/180,001 bindings at 20k/60k emitted records; the new one has three at both points. Old forced-GC heap growth is about 5.5 MiB; the new recorded growth is 0–6 KiB across engines and position units. Live snapshots, scalar data and memo keys retain what their contracts require; this change bounds binding history, not arbitrary user values or nesting.
• Six interleaved before/after runs against f259b37, Apple M3 Max, Go 1.27.1, -benchtime=200ms, recognition with code points: reading limit while repeatedly binding i takes 169–171 → 0.65–0.82 ms at 16,000 characters. The new timings grow linearly from 2k through 16k, whereas the old timings grow quadratically. Nearest-variable time drops 13–14%, and equal-value reuse reduces allocation from about 0.77 MB/16,000 allocations to 2.5–2.8 KB/9–15 allocations. Alternating constant names also avoid copying.
• Changed values cost more allocation: incrementing two alternating counters while reading limit takes 345–350 → 2.34–2.59 ms at 16k, but allocation rises 1.86 → 3.33 MB (+78–79%) and about 63,600 → 95,500 allocations (+50%). A failed choice arm that changes both counters before retrying takes 403–431 → 4.49–4.94 ms at 16k, with 3.65 → 6.65 MB (+82%) and about 127,100 → 191,000 allocations (+50%). Copying the prefix buys immutable rollback and bounded lookup/retention. Many distinct first bindings also pay for the hint map: 16 names take 0.46 → 1.27 µs/2.3 → 4.2 KB, 128 take 2.22 → 9.06 µs/7.7 → 21.5 KB, and 1,024 take 16.9 → 71.6 µs/50.7 → 160.1 KB. These synthetic startup costs are accepted for the retention fix; they are not general speedups.
• The existing Outline batch workload stays within 2.7% in median time for parsing/recognition on every engine, generated Node Go and generated typed Go. The hint map adds about two allocations; bytes increase by at most 0.03% in this sample. TS live-state and parity tests cover semantics; no TS speedup is measured. The full benchmark report is refreshed at the stream-memory checkpoint.
• Six pairs compare f259b37 (runtime and generated benchmark fixtures restored with a Go overlay) with the unique-binding change in ec2f903. Restore runtime.go, eval.go, vm.go and internal/bench/gen/*/parser.go; an unused prepend-only bind helper lets new snapshot tests compile on the baseline, whose binding paths and parser layout remain unchanged. The rollback workloads and heap regressions are in the benchmark/test code. Reproduce with go test ./internal/engine -run '^$' -bench 'Benchmark(EnvironmentBindings|DistinctBindings)$' -benchmem. Tests check saved snapshots against independent value maps, choice/rule scope, positive/negative lookahead, streams and actual generated Go/TS rule state.