feat: bulk forget and negative concept filters for recall

[CalebisGross]

Bulk forget

forget currently accepts a single memory ID. When cleaning up test data or archiving a batch of related memories, this requires N separate tool calls.

Proposed: Accept an array of IDs:

{"memory_ids": ["id1", "id2", "id3"]}

Keep the existing single-ID interface for backwards compatibility.

Negative concept filters

Recall currently supports positive concept filtering (concepts: ["retrieval"]), but there's no way to exclude concepts. When searching for retrieval-related memories, training results often surface because they share concepts.

Proposed: Add exclude_concepts parameter:

{"query": "retrieval ranking", "concepts": ["retrieval"], "exclude_concepts": ["training", "sweep"]}

Filter out any memory whose concepts overlap with the exclude list before ranking.

Why together

Both are MCP tool ergonomics improvements that touch internal/mcp/server.go handler logic and would be natural to ship in one PR.

[CalebisGross]

Fixed in PR #312 — forget now accepts memory_ids array for bulk archival (backwards-compatible with single memory_id). recall now accepts exclude_concepts array applied in both the retrieval agent applyFilters and the concept-only search path.