feat: vector similarity-based behavioral pattern detection

[BYK]

Summary

Adds pattern echo detection to the distillation pipeline. When a new distillation segment has high embedding similarity to segments from previous sessions, it signals repeated user behavior. The system extracts the common pattern via LLM and creates a preference knowledge entry.

How it works

1. After each gen-0 distillation segment is created, its embedding is compared against all previous segments for the same project via vectorSearchAllDistillations()
2. Segments with cosine similarity >= 0.78 from different sessions are counted as "echoes"
3. When 2+ echoes from distinct sessions are found, the pattern extraction LLM synthesizes the common behavioral pattern
4. A preference entry is created at confidence 0.8 (auto-extracted, below user-stated 1.0)
5. ltm.create() dedup guard prevents duplicate entries
6. Rate limited to 1 extraction per session per 10 minutes

Detected patterns in eval

• "Always use centralized error handling with try/catch and semantic HTTP codes"
• "Never use ORMs — always use raw parameterized SQL via pg"
• "Always use const by default; only use let when reassigned"
• "Always follow a strict post-implementation checklist for API endpoints"

Limitations

Pattern echoes detect topically similar behaviors (same phrasing across sessions). They miss abstract behavioral patterns where the surface text varies — e.g., "always asks for tests after implementation" when the tests are for REST endpoints in one session and React components in another. These require higher-level temporal/sequential reasoning.

Files

• packages/core/src/pattern-echo.ts (NEW) — detection logic, ~170 lines
• packages/core/src/prompt.ts — PATTERN_ECHO_SYSTEM prompt + patternEchoUser()
• packages/core/src/distillation.ts — hook into distillSegment(), await when urgent