feat(gateway): add smart_model_routing for auto-complexity classification

[davidgut1982]

What

Adds a smart_model_routing gateway-level classification step that automatically routes short, simple messages to a configurable cheap model before agent dispatch — without requiring any user action or session state change.

How it works (gateway/run.py):

GatewayRunner._apply_smart_routing(message, model, runtime_kwargs, smart_cfg) inspects the raw incoming message text:

• If the text is shorter than max_simple_chars (default 200) and max_simple_words (default 40)
• And contains none of complexity_keywords (e.g. "implement", "debug", "refactor", "diagnose", "architect")
• Then runtime_kwargs["model"] is replaced with cheap_model before the agent starts

The gate is applied in run_sync() after _resolve_session_agent_runtime and is skipped when:

• The session has an active /model override (_session_model_overrides entry present)
• The source platform is Platform.LOCAL (CLI sessions bypass smart routing)

Configuration (hermes_cli/config.py):

smart_model_routing:
  enabled: false          # opt-in; must explicitly set true
  cheap_model: deepseek/deepseek-v4-flash
  max_simple_chars: 200
  max_simple_words: 40
  complexity_keywords:
    - implement
    - debug
    - refactor
    - diagnose
    - migrate
    - architect
    - explain
    - "why does"
    - "how does"
    - broken
    - failing
  simple_keywords:        # informational only; short + no complexity_keyword → cheap
    - status
    - show
    - check
    - list
    - restart
    - "what is"
    - ping
    - health

All provider routing, API keys, and base URLs flow through unchanged; only the model slug in runtime_kwargs is swapped.

Why

Operators running Hermes on a fixed budget or with high message volume pay the same cost for "ping" as for "implement OAuth2". Smart routing cuts cost on trivial turns without changing the agent's behavior, memory, or session continuity.

Tests

pytest tests/gateway/test_smart_model_routing.py -v

6 tests covering: simple message routed to cheap model, complexity keyword blocks downgrade, message-too-long blocks downgrade, manual /model override bypasses smart routing, enabled: false passthrough, and platform-LOCAL bypass. All pass.

Platforms tested

Linux (CT/LXC environment, Python 3.13).

[combatsheep]

Smart routing is well-contained: the override and Platform.LOCAL bypasses stay intact, and the classifier only rewrites runtime kwargs for short, non-complex messages. Approving.

[teknium1]

We do not want this