KB and MC gate their diagnostic reads behind full operability, so an embedding outage blinds the surface needed to diagnose it

Context

During a live external-plane embedding outage on 2026-08-13, Knowledge Base refused list_documents and get_deployment_state_snapshot because aggregate health included a failed embedding probe. Storage and the orchestrator deployment-state bridge remained readable, so the failed dependency did not prevent either requested read.

The same incident showed the correct opposite case: semantic queries genuinely need the embedding provider and must continue to refuse while it is unavailable.

The Problem

Knowledge Base already has a per-tool health-gate exemption seam, Server.getHealthExemptTools(), but three diagnostic reads are missing from it:

• list_documents reads the tenant-scoped Chroma collection through DocumentService.listDocuments. It does not embed.
• get_deployment_state_snapshot reads the orchestrator snapshot file.
• inspect_deployment dispatches to the same snapshot-file reader.

As a result, an embedding-only outage blinds the existing surfaces needed to determine corpus coverage and deployment state.

Memory Core is not part of the remaining implementation. #14124 and PR #14162 already repaired the equivalent catch-22 there by extending its existing exemption list; current Memory Core keeps embedding queries gated and its non-embedding diagnostics available.

Fix Shape

Extend the Knowledge Base Server.getHealthExemptTools() list with:

• list_documents
• get_deployment_state_snapshot
• inspect_deployment

Reuse the existing BaseServer dispatch policy. Do not add a second dependency registry, a new MCP tool, or a new partial-response schema.

The existing tool output contracts stay unchanged. If storage or the snapshot file itself is unavailable, the handler still returns its specific failure; the exemption only prevents an unrelated embedding-health failure from blocking dispatch.

Acceptance Criteria

• With the embedding health gate failing, list_documents dispatches and returns its existing result when storage is live.
• Under the same condition, get_deployment_state_snapshot and inspect_deployment dispatch through their existing file-read helper.
• query_documents and ask_knowledge_base still refuse before dispatch because they genuinely require embeddings.
• The regression fixture proves both halves at the BaseServer request boundary: diagnostic reads bypass; an embedding query remains gated.
• No existing MCP response schema changes.

Out of Scope

• Repairing the embedding provider or provider lane.
• Memory Core health-gate changes; its equivalent diagnostic exemption already exists.
• General classification of every read-like KB tool.
• New MCP tools or a generic dependency-policy framework.

Avoided Traps

• Do not globally weaken health status. The embedding outage remains real and visible.
• Do not exempt semantic queries that embed at request time.
• Do not annotate successful diagnostic payloads with a new ad hoc partial shape; unchanged output plus honest handler failures keeps the contract stable.

Evidence and Precedent

• #14124 and PR #14162: established Memory Core exemption pattern.
• Current KB Server.getHealthExemptTools(): existing per-server policy seam.
• Current KB toolService: both deployment tools share readDeploymentInspection; list_documents maps to DocumentService.listDocuments.
• Parent epic: #17072; independent epic review green at https://github.com/neomjs/neo/issues/17072#issuecomment-5287150119

Authored-by: @neo-opus-vega. Intake correction and implementation by @neo-gpt.