reliability: memory durability — crash-recovery + concurrent-write tests

[matedev01]

Goal: the memory subsystem is durable and concurrency-safe — the baseline portability must not weaken.

Sub-steps

• Crash-recovery test on the WAL/pruning path (feat(storage): add automatic db pruning and storage health endpoint #516/perf(memory): batch prune_decayed deletes instead of one DELETE per row #542).
• Concurrent-write test.
• Dispatch arg-validation for memory tools ([bug] tools/dispatch: memory_forget and memory_store accept missing required args with soft success — complete M1 typed-tool enforcement #416).

M2 · Workstream D.

[matedev01]

All sub-steps are complete:

• Crash-recovery test on the WAL/pruning path — Add crash-recovery and concurrent-write durability tests for the memory subsystem #576 (wal_pruning_path_recovers_after_unclean_shutdown).
• Concurrent-write test — Add crash-recovery and concurrent-write durability tests for the memory subsystem #576 (concurrent_writers_serialize_without_corruption).
• Dispatch arg-validation for memory tools ([bug] tools/dispatch: memory_forget and memory_store accept missing required args with soft success — complete M1 typed-tool enforcement #416) — test(tools): close #416 memory_forget/memory_store schema enforcement #581.

On top of that, the memory subsystem now runs its SQLite off the reactor thread (#606), so a memory operation can no longer stall other in-flight sessions. Closing — D4 memory durability is done.