feat: Research vault recall — check prior research before spawning new work

[qozle]

Summary

The research vault (MEMORY/RESEARCH/) persists research outputs across sessions, but no workflow proactively checks it when relevant topics arise. Research agents are spawned fresh every time, even when prior deep investigations on the same or related topics already exist in the vault.

This is distinct from #844 (which covers the capture mechanism — getting research to persist to disk). This issue covers recall — proactively checking existing vault content before spawning new research or analytical work.

Problem

• User researches "AI trading firms" via Deep Investigation → vault populated with landscape, entities, profiles
• User later asks to research "quantitative trading strategies" → 9 agents spawned from scratch with no awareness of the prior AI trading research
• User runs WorldThreatModelHarness to test an investment thesis → 11 world model agents reason from training data alone, despite relevant prior research existing on disk
• Algorithm OBSERVE phase generates ISC criteria without knowing what the user has already investigated

Proposed Solution

Add a lightweight vault relevance check at three integration points:

1. Algorithm OBSERVE phase

• After environment discovery, before ISC generation
• Scan MEMORY/RESEARCH/ file names and frontmatter titles
• Match against current request keywords
• If relevant: note in ISC context as available resource for subsequent phases
• If none: note "no relevant prior research" and move on

2. Research skill workflows (Quick, Standard, Extensive, DeepInvestigation)

• Before spawning research agents
• Same scan as above
• If relevant: include summary in agent briefs ("Prior research found: [summary]. Build on these findings.")
• If comprehensive and recent (<30 days): suggest extending rather than starting fresh
• DeepInvestigation: cross-reference broader vault during Step 0 iteration detection

3. WorldThreatModelHarness TestIdea

• Before dispatching to world model agents
• If relevant vault research exists: include summary as grounding context in agent briefs
• Gives world model agents concrete prior findings instead of relying solely on training knowledge

Design Constraints

• Lightweight: ls + at most 2-3 frontmatter reads per check. No spawned agents for the check itself.
• Graceful degradation: Empty vault = skip the check, no errors.
• No new files: Surgical insertions into existing workflow steps.
• Keyword matching: File names and frontmatter titles only — not full-text search.

Files Changed

• PAI/Algorithm/v3.7.0.md — OBSERVE phase vault check step
• skills/Research/Workflows/QuickResearch.md — Step 0.5 vault check
• skills/Research/Workflows/StandardResearch.md — Step 0.5 vault check
• skills/Research/Workflows/ExtensiveResearch.md — Step 0 vault check
• skills/Research/Workflows/DeepInvestigation.md — Step 0 cross-vault check
• skills/Thinking/WorldThreatModelHarness/Workflows/TestIdea.md — Step 1.5 vault check

Related

• Potential Bug: MEMORY/RESEARCH/ capture mechanism. #844 — Research capture mechanism (complementary: capture writes to vault, recall reads from it)