67. Retire Memo Entries When Their Lookup Slots Are Removed

• Pruning a stream's lookup slots or replacing a Document key left the old entry's result reachable from its allocation slab. Entries now clear their node, environment, error and expectation references when retired, and reuse their storage. Pruning preserves a seed still owned by an active left-recursion frame; replacing a seed retires it after that frame installs its next best.
• Regression measurements on every engine backend and both position units: 20,000 to 60,000 discarded abc,abc\n records retained 42 MiB with memoized words, or 135–137 MiB with left-recursive words. With retirement the sampled growth is 34–137 KiB. The left-recursive case holds two memo chunks and 257 allocated entries at 20k, 60k and 200k records; the remaining heap variation follows the 1,024-position pruning interval. Replacing the first character of a 10 KB Document and parsing after each edit retained 13 MiB between edits 20 and 300; the corresponding samples now range from 30 KiB lower to 9 KiB higher.
• Effect (median of six interleaved before/after runs, Apple M3 Max, Go 1.27.1, -benchtime=200ms, code points): BenchmarkMemoLifecycle, 10,000 records, allocates 11.0–11.3 → 7.8–8.1 MB for memoized words (−28% to −29%), and 35.9–36.6 → 23.1–23.9 MB for left-recursive words (−35% to −36%). Time changes are −1% to −6% and −5% to −9%, respectively. The existing Arith_LeftRec batch workload allocates 4–5% fewer bytes for parsing and 19–21% fewer for recognition; time is within 4% on every backend/unit. Generated runtimes are unchanged. These are focused comparisons against f3a2c96 with an overlay of its memo allocator, compared with the retirement change in 9353e09; this was not a refresh of the full benchmark report.
• Reproduce the focused stream cases with go test ./internal/engine -run '^$' -bench '^BenchmarkMemoLifecycle$' -benchmem. Memory regressions are TestMemoizedStreamMemoryIsBounded and TestDocumentReplacedMemoMemoryIsBounded.