Budget-aware multi-model routing with per-task cost tracking

Start in the affected terraphim_service, terraphim_types, terraphim_multi_agent, and terraphim_mcp_server crates, then review the event-sourcing dependency in #597 and the existing terraphim-llm-proxy token headers. Done means task budgets meter costs, enforce limits, fall back through model chains, persist cost events, and expose the current and historical MCP budget resources.

Summary

Add per-task cost tracking with fallback chains and hard spending limits to LLM routing. Each agent operation gets a budget allocation; if the primary model is too expensive or unavailable, fall back to cheaper alternatives automatically.

Motivation

Inspired by Ouroboros per-round cost tracking with per-worker budget allocation. Current state:

• terraphim-llm-proxy routes between multiple models but has no per-task budget tracking
• genai (terraphim fork) handles LLM calls but does not meter usage or cost
• The multi-model judge pattern (quick -> deep -> tiebreaker) needs budget awareness: only escalate to expensive models when cheaper ones disagree
• No spending limits exist -- a runaway agent loop could consume unbounded API credits

Implementation Plan

1. Budget Struct

pub struct TaskBudget {
    pub max_cost_usd: f64,
    pub model_chain: Vec<ModelConfig>,  // ordered by preference (cheapest first)
    pub spent_usd: f64,
    pub request_count: u32,
    pub created_at: jiff::Timestamp,
}

pub struct ModelConfig {
    pub model_id: String,
    pub cost_per_1k_input: f64,
    pub cost_per_1k_output: f64,
    pub max_latency_ms: u64,
    pub fallback_on: Vec<FallbackReason>,  // Timeout, RateLimit, BudgetExceeded, EmptyResponse
}

2. Cost Tracking in terraphim_service

• After each LLM call, compute cost from token counts + model pricing
• Accumulate in TaskBudget.spent_usd
• If spent_usd >= max_cost_usd, reject further calls for this task
• Log cost events to event store (#597)

3. Fallback Chain Logic

• Try models in model_chain order
• On failure (timeout, rate limit, empty response, budget exceeded for this model tier), try next model
• If all models exhausted, return error with cost summary

4. Integration with terraphim-llm-proxy

• Proxy already reports token counts in responses
• Add cost metadata to proxy response headers (cost_usd, model_used, tokens_in, tokens_out)
• Agent-side budget tracking consumes these headers

5. MCP Resource Exposure

• Expose /budget/current as MCP resource showing active task budgets
• Expose /budget/history for cost analysis over time

Affected Crates

• terraphim_service (primary -- add budget tracking to LLM routing)
• terraphim_types (add TaskBudget, ModelConfig types)
• terraphim_multi_agent (wire budget into agent task execution)
• terraphim_mcp_server (expose budget resources)

Dependencies

• #597 Event sourcing (for cost event persistence)

Estimated Effort

~4 hours (mostly plumbing between terraphim_service and proxy response headers)

Part of

Epic #595