feat: improve meta-request detection to prevent distilling title/summary conversations

[BYK]

Summary

• Replace isTitleOrSummaryRequest() with isMetaRequest() using a two-layer detection approach to prevent meta conversations (title generation, summaries, categorization) from being stored and distilled
• Add chat.headers hook to OpenCode plugin injecting x-lore-agent header for authoritative agent identification
• Improve heuristic scoring with maxTokens and keyword signals alongside existing structural checks

Details

Problem: The gateway's meta-request filter used a narrow structural heuristic (tools ≤ 2, messages ≤ 2, system < 500 chars) that could miss title/summary requests with slightly different shapes, causing them to be stored in temporal and distilled into session history.

Layer 1 — Explicit header: The OpenCode plugin SDK provides the agent name (e.g., "coder", "title") via the chat.headers hook. We now inject it as x-lore-agent, giving the gateway an authoritative signal. Known meta agents (title, summary, categorize, etc.) are always passthrough; known primary agents (coder, code) are never meta.

Layer 2 — Improved heuristics: For non-OpenCode clients (Claude Code, Pi, generic), we replace the rigid boolean gate with weighted scoring:

Signal	Score	Rationale
tools ≤ 2	+3	Real agents have 5+ tools
messages ≤ 2	+3	Meta requests are single-shot
system < 500 chars	+2	Real prompts are 2K–50K
maxTokens ≤ 300	+3	Title gen uses tiny output budgets
Meta keywords in short system prompt	+2	Bonus only, never standalone

Threshold: ≥ 8 (preserves backward compat — the old 3-check AND scored 3+3+2 = 8).

Files changed:

• packages/opencode/src/index.ts — chat.headers hook
• packages/gateway/src/compaction.ts — isMetaRequest() with scoring
• packages/gateway/src/pipeline.ts — updated import, call site, observability log
• Tests and comments updated across 6 files