[fred-doctorina]

Why

The context-service search seam (__fredContextServiceSearch) is bounded by an abort timeout so a slow/cold context-service call can never stall a turn. That budget was hard-coded at 800ms.

Measured on the live deploy: a cold query-embedding round-trip through context-service takes ~700-760ms (OpenAI embedding warm-up + pg pool), while warm calls are ~150ms. At 800ms the first (cold) query per idle period is aborted and the seam silently falls through to the dead native memory path -> provider:none / backend:builtin / 0 hits. This is the root cause of intermittent "memory search returns nothing / index looks disabled" reports.

What

• Extract resolveContextServiceTimeoutMs() and use it for the seam AbortSignal.timeout(...).
• Default raised 800ms -> 2500ms (covers a cold call with margin).
• Configurable via FRED_CONTEXT_SERVICE_TIMEOUT_MS; clamped to [100, 30000]ms; invalid/blank -> default.
• Unit tests for default / parse / clamp / invalid.

Notes

• No behavior change on success path other than the larger budget.
• Base branch is doctorina/base-v2026.6.6 (the branch the gateway image is built from; main does not carry the seam).
• Follow-up (separate, optional): boot-warm the embedding connection in services/context-service so even the first post-redeploy query is warm.