DummyLM should share a trait with LM

[darinkishore]

Problem

LM and DummyLM are structurally separate types with duplicated APIs:

• Both have call(), inspect_history(), cache_handler, temperature, max_tokens
• But they don't share a trait — code that wants to be generic over "any LM" can't be

DSPy Python has BaseLM as the abstract base class that both real and mock implementations extend.

Proposal

Extract a shared trait:

#[async_trait]
pub trait LanguageModel: Send + Sync {
    async fn call(&self, messages: Chat, tools: Vec<Arc<dyn ToolDyn>>) -> Result<LMResponse>;
    async fn inspect_history(&self, n: usize) -> Vec<HistoryEntry>;
    fn model(&self) -> &str;
}

LM and DummyLM both impl this. GLOBAL_SETTINGS stores Arc<dyn LanguageModel>. Modules reference the trait, not the concrete type.

Related

• LM layer: per-module LM, config merging, retry/backoff #69 (LM layer: per-module LM)

[darinkishore]

Codebase audit — LM vs DummyLM duplication

LM (core/lm/mod.rs):

• call(&self, messages, tools) → Result<LMResponse>
• execute_tool_loop(...) → tool iteration
• inspect_history(n) → reads from cache
• Fields: model, temperature, max_tokens, cache, cache_handler, client, api_key, base_url, max_tool_iterations

DummyLM (core/lm/mod.rs:554+):

• call(&self, prompt) → returns canned responses
• inspect_history(n) → reads from its own cache
• Fields: api_key, base_url, temperature, max_tokens, cache, responses

Settings (core/settings.rs):

pub struct Settings {
    pub lm: Arc<LM>,  // concrete type, not trait
    pub adapter: Arc<dyn Adapter>,
}

Note adapter is already dyn Adapter (trait object) but lm is Arc<LM> (concrete). Inconsistent.

Proposed trait:

#[async_trait]
pub trait LanguageModel: Send + Sync {
    async fn call(&self, messages: Chat, tools: Vec<Arc<dyn ToolDyn>>) -> Result<LMResponse>;
    async fn inspect_history(&self, n: usize) -> Vec<HistoryEntry>;
    fn model(&self) -> &str;
    fn temperature(&self) -> f32;
    fn max_tokens(&self) -> u32;
}

Migration:

1. Define LanguageModel trait in core/lm/mod.rs
2. Impl for both LM and DummyLM
3. Change Settings.lm to Arc<dyn LanguageModel>
4. Change get_lm() to return Arc<dyn LanguageModel>
5. Update Predict and anywhere else that uses Arc<LM> directly

This unblocks: #69 (per-module LM — modules store Option<Arc<dyn LanguageModel>>), better testing (inject DummyLM anywhere LM is expected), and third-party LM implementations.