Implement TAGGED_CONCEPT auto-emit via SemanticGraphExtractor on message bodies

Context: Deferral from the Phase 3 Mailbox A2A integration (#10146, #10147). The core MailboxService currently allows clients to manually specify taggedConcepts array, but the automatic graph emission using the SemanticGraphExtractor on message bodies was deferred.

The Problem: Manual tagging is prone to omission and does not leverage the LLM capabilities of the swarm to automatically build knowledge density around specific concepts as agents converse.

The Architectural Reality: This is an extractor-layer concern. The SemanticGraphExtractor needs to be wired into the MailboxService's addMessage flow (or operate as a decoupled listener on the graph) to auto-extract concepts from bodyText and emit TAGGED_CONCEPT edges linking the Message node to Concept nodes.

The Fix:

• Implement automatic concept extraction from message bodies using SemanticGraphExtractor.
• Ensure extraction handles failure gracefully without blocking message delivery.

Acceptance Criteria:

• When a message is sent, its body is processed by the Gemma4 extractor (or equivalent).
• TAGGED_CONCEPT edges are automatically created linking the message to the extracted concepts.

Out of Scope: Other edge types (already handled by Phase 3).

Origin Session ID: ada4aecb-60fc-4b25-a76b-cb92600835a2