5. Memo Table with per-Position Chains (Cf77e4f)

• The memo was a Go map keyed by (rule, position, level): hashing, map growth and one allocation per entry.
• Entries now hang off a slice indexed by position and are allocated in slabs of 256. Stream pruning drops a prefix of the slice. Kept expectation sets are copied into fixed-size chunks, and a repeat of the previous set is shared.
• Effect: JSON ~130 → ~85 ms, minilang ~103 → ~65 ms; about a third of parse time.