feat(chat): tiered L0-L3 context wake-up to replace unconditional prompt injection #5066

Problem

AutoBot currently loads context into agent sessions via unconditional snippet injection (e.g. `CAUSAL_REASONING_SNIPPET` — see #4811) and full chat-history recall. This wastes tokens and degrades signal-to-noise: every agent gets the same preamble regardless of what the session is about.

There is no token-budgeted, tiered wake-up strategy. The agent either has too little context (cold start) or too much (session with accumulated history).

Proposed change

Introduce a 4-tier wake-up stack in `autobot-backend/chat_history/layers.py`:

• L0 Identity (~100 tok, always loaded) — who this agent is, role, owner. From a small config file / identity table.
• L1 Essential Story (~500-800 tok, always loaded) — auto-generated from highest-weight / most-recent memory-graph entities and KB facts for the current user/session.
• L2 On-Demand (~200-500 tok each) — loaded when a topic/entity is mentioned in the user message; scoped pull from memory graph and KB.
• L3 Deep Search (unlimited) — existing hybrid search + rerank path, invoked only for explicit retrieval-heavy queries.

Target: wake-up cost ≤ 900 tokens for L0+L1, leaving ≥95% of the context window for actual conversation.

Acceptance criteria

• `Layer0/Layer1/Layer2/Layer3` classes with a common `render()` + `token_estimate()` interface
• Wired into chat_workflow/manager.py session bootstrap
• L0+L1 combined cost measured and asserted < 900 tokens in a test
• L2 fires based on entity/topic detection in the user message
• Feature flag to A/B against current unconditional-snippet path
• Replaces discovery(agents): CAUSAL_REASONING_SNIPPET unconditionally injected into every IntelligentAgent LLM prompt #4811 unconditional injection (link and close once shipped)

Related

• discovery(agents): CAUSAL_REASONING_SNIPPET unconditionally injected into every IntelligentAgent LLM prompt #4811 (CAUSAL_REASONING_SNIPPET unconditional injection)