fix(openviking): synchronous prefetch with current turn's query

[wgd753]

Problem

The OpenViking memory provider's prefetch() method returns stale results from a background thread that fired at the end of the previous turn. The current turn's query parameter is completely ignored, meaning automatic memory context is always about the wrong topic.

First turn gets zero context. Every subsequent turn gets stale context from the previous topic.

Root Cause

plugins/memory/openviking/__init__.py — prefetch() currently ignores its query parameter and returns whatever the background thread cached from the previous turn's topic.

Fix

prefetch() now performs a synchronous POST /api/v1/search/find using the current turn's query parameter — matching Claude Code's plugin behavior.

1. Uses the actual query — searches for the current topic, not the previous one
2. Score threshold — minimum _PREFETCH_MIN_SCORE = 0.35, filters noise
3. Timeout guard — _PREFETCH_TIMEOUT = 1.5s prevents slow OV from blocking the turn loop
4. Stale cache drained — previous turn's background-thread result is discarded

queue_prefetch() is unchanged — it still fires background searches at turn-end for pre-warming, but prefetch() no longer consumes its output.

Related

• Ref: All agent implementations are not the same: Hermes Agent integration can't be compared to the Claude Code and OpenClaw plugins volcengine/OpenViking#2253 (root cause analysis)
• Closes: feat(memory): Allow synchronous recall for current turn as option for honcho and hindsight #5820
• Contrast: PR feat(memory): add sync_recall option for current-turn relevance #5838 (sync_recall manager-layer approach — not merged; this PR fixes it at the provider level, similar to the merged Honcho fix fix(honcho): pass user_message as search_query for topic-relevant context (salvage #17021) #20178)

[alt-glitch]

Duplicate of #33260 — same fix: making OpenViking prefetch() perform a synchronous POST /api/v1/search/find with the current turn's query instead of returning stale background-thread results, with the same score threshold (0.35) and a timeout guard. #33260 (open, filed 2026-05-27) is the canonical version. Related: #32444 (broader OpenViking runtime-reliability PR covering the same code path).

[wgd753]

Agreed, this is a duplicate of #33260 — same core fix (sync prefetch with current query, score threshold 0.35, drain stale cache).

One difference our version has that #33260 doesn't: a _PREFETCH_TIMEOUT = 1.5s guard to prevent a slow/unreachable OV server from blocking the agent turn loop. Would be great if #33260 could pick that up.