[Bug] recallScoreThreshold bypassed in pickMemoriesForInjection function

Description

The recallScoreThreshold configuration is being bypassed during memory recall, causing low-scoring memories (below threshold) to be injected into context.

Evidence

• Config: recallScoreThreshold: 0.25
• Observed: Retrieved memories include scores as low as 0.11 (below threshold)
• Expected: All returned memories should have score ≥ 0.25

Root Cause Analysis

The issue is in memory-ranking.ts, specifically in the pickMemoriesForInjection function.

Flow:

1. postProcessMemories correctly filters memories by threshold (line 49):

   if (clampScore(item.score) < options.scoreThreshold) {
     continue; // ✓ Correct filtering
   }

2. However, pickMemoriesForInjection (line 221) bypasses this filtering when trying to fill up to the limit:

   const leaves = deduped.filter((item) => isLeafLikeMemory(item));
   if (leaves.length >= limit) {
     return leaves.slice(0, limit);  // ✗ No threshold check
   }
   
   const picked = [...leaves];
   const used = new Set(leaves.map((item) => item.uri));
   for (const item of deduped) {
     if (picked.length >= limit) {
       break;
     }
     if (used.has(item.uri)) {
       continue;
     }
     picked.push(item);  // ✗ Adds items without threshold check!
   }
   return picked;

Problem:

When there aren't enough "leaf-like" memories (level === 2), the function supplements from the general deduped array without re-checking the score threshold. This allows low-scoring memories to pass through.

Proposed Fix

Add threshold validation in pickMemoriesForInjection:

for (const item of deduped) {
  if (picked.length >= limit) {
    break;
  }
  if (used.has(item.uri)) {
    continue;
  }
  // ADD: Re-check threshold before adding
  if (clampScore(item.score) < options.scoreThreshold) {
    continue;
  }
  picked.push(item);
}

Note: The function signature would need to accept scoreThreshold as a parameter, or use a default value.

Impact

• Low-quality/irrelevant memories are being injected into agent context
• Wastes token budget on low-scoring results
• Reduces overall recall quality

Environment

• OpenViking: v0.2.15
• Plugin installation: path-based (/Users/neoshi/.openviking/plugins/openviking)
• Config mode: remote

Additional Context

This was discovered during investigation of recall quality after disabling rerank. Statistics from 14 queries:

• Average score: 0.2827
• Score range: 0.1100 - 0.7222 (minimum below 0.25 threshold)
• Zero-result rate: 0%