Resolves #13950

LM Studio loaded-model readiness now treats non-numeric parallel telemetry as unobservable instead of an unsatisfiable gate, while still repairing numeric chat parallel mismatches. The readiness helper also exposes a serialized, size-capped OpenAI-compatible embedding-serving canary that reports only diagnostics and vector length, never vector bodies.

Evidence: L2 (focused unit fixtures plus static diff/syntax checks) -> L2 required (the close target explicitly forbids live large-input stress probes). Residual: operator-gated local orchestrator restart smoke after merge.

Deltas from ticket

The embedding canary is implemented as an opt-in readiness seam, not an automatic orchestrator startup request. That is intentional: the incident evidence showed direct provider probes can contend with KB sync or maintenance work. Callers must pass the canary and can skip/degrade it through the load gate.

Existing TextEmbeddingService loaded-context enforcement remains the hard stop for provider-side embedding truncation. This PR does not change embedding context defaults or add any env/config fallback.

Test Evidence

• npm run test-unit -- test/playwright/unit/ai/scripts/runners/runSandman.spec.mjs -> 62 passed.
• git diff --check -> passed.
• node --check ai/services/graph/providerReadinessHelper.mjs -> passed.
• Commit hook passed: whitespace, shorthand, AiConfig test-mutation, JSDoc types, ticket archaeology, and block-alignment checks.

Post-Merge Validation

• Pull dev, run node ./ai/scripts/setup/initServerConfigs.mjs --migrate-config, then restart the harness/orchestrator.
• With models idle, confirm LM Studio keeps chat + embedding resident and no reload loop is caused by embedding parallel being unreported.
• Run any embedding-serving canary only when no KB sync/heavy maintenance is active.

Commits

• a0f0810556 - fix(ai): make lms embedding readiness load-aware (#13950)

Authored by Euclid (GPT-5, Codex Desktop). Session cd2b88d7-8134-4ef8-a10f-0b1e80c03db4.