Preserve updated Q after cache eviction

[Frankie-Xu]

MemoryService keeps both a fast _q_cache and a _mem_cache of memory objects. After a successful Q update, FIFO eviction can remove an ID from _q_cache; retrieval then falls back to that memory object's stale metadata.q_value and can silently lose the learned value.

Synchronize cached memory metadata after successful non-None single or batch updates. If the metadata object rejects assignment, invalidate that _mem_cache entry so the next retrieval uses the existing storage read path instead of reusing a stale fallback. This preserves the Q-learning formula, FIFO policy, retrieval scoring, and existing error contract without adding cross-process or concurrency behavior.

Validation:

• On main c1b322c, the original single/batch eviction regressions failed (learned -0.5/0.5 fell back to 0.0).
• A read-only metadata regression then failed on the first PR17 revision because the stale entry remained cached.
• At head 4a2d77c, five unittest cases pass; the same five pass under pytest and from a clean git archive export.
• Coverage includes single and batch updates, FIFO eviction, negative and zero Q values, persistence failures without cache/metadata pollution, q-cache snapshot round-trip, rejected metadata assignment, and fresh storage rehydration.
• Tests execute the real MemoryService retrieval/update and QValueUpdater paths with only MemOS import types, in-memory persistence, fixed embeddings, and a contract-shaped read-only metadata double substituted. MemoryOS==1.0.0 is declared but not installed locally, so no live integration is claimed.
• compileall and git diff --check pass.

This is independent of PR #14's finite-value validation and PR #16's single-update Q-cache publication.