Mailbox: optional Chroma semantic layer for "find related messages"

Context

Parent #10139 phase 4 (optional). SQLite FTS5 + concept edges cover ~90% of mailbox search. This ticket adds a Chroma neo-mailbox collection for semantic body search as specialized query layer. Phase-gated: ship after phases 1-3 deliver value, only if friction surfaces.

The Problem

Some use cases benefit from semantic retrieval over messages: "find prior handoffs where I discussed mailbox design" without exact keywords, or "find replies to messages similar in intent to X." Graph concept edges cover part, FTS5 another part, but semantic-over-body fills the remaining gap.

The Architectural Reality

• Chroma already runs for Memory Core. Adding a neo-mailbox collection is infrastructure-zero.
• Messages are lower cardinality than memories (10s-100s/month vs thousands) — embedding cost bounded.
• Primary lookup stays graph-primary; Chroma is sidecar for semantic-over-body.

The Fix

• New Chroma collection neo-mailbox; embedding pipeline identical to existing memory pipeline
• Tool findRelatedMessages({query, limit, filters}) — top-k with tenant scope at result-filter layer
• Sync: messages embed on addMessage; retractions clear embedding
• NOT a primary lookup path

Acceptance Criteria

• Chroma neo-mailbox collection provisioned
• Embedding hook in addMessage
• findRelatedMessages tool; tenant-scoped results
• Tool description differentiates: "semantic similarity" vs listMessages "filter + sort"
• Retraction clears embedding

Out of Scope

• Replacing graph-primary access paths with Chroma (category separation must hold)
• Cross-tenant semantic search (tenant-scoped, same as memories)

Avoided Traps

• Chroma as primary lookup. Rejected. Relationships dominate mailbox access; semantic is auxiliary. Chroma-primary would tank hot paths.

Related

• Parent: #10139
• Depends on: Message schema sub; phases 1-3 should ship first to observe if friction is real
• Files: ai/mcp/server/memory-core/ semantic handler (new), Chroma schema migration

Origin Session ID: 71dc3cd8-d39d-48e1-ac62-e240ca67d1a5