Persist CLI cost and token telemetry in debate records

[simonhsze]

What

Expose per-run token and cost telemetry in the CLI artifacts and governance history, instead of keeping cost accounting only inside adapters / agent traces.

Motivation

The project already computes input_tokens, output_tokens, and cost_usd at the adapter layer, and BaseAgent accumulates per-agent totals through CostGuard. But the normal CLI flow does not persist that data into workspace/governance_history.json or surface it in debate artifacts.

That creates an observability gap:

• cost exists internally but is not first-class in the user-facing run record
• mixed-model runs can become materially more expensive without the CLI making that obvious
• ac score and downstream analysis cannot reason about debate quality vs cost tradeoffs

This is separate from issue #8, which is about whether soft-limit warnings belong in the runtime path.

Scope

• persist total input/output tokens and total cost for each CLI run
• optionally break out cost by role: analyst / critic / judge
• include cost metadata in saved debate artifacts and/or CLI summary output
• decide whether governance history should store raw telemetry, aggregates, or both
• add tests covering real adapters and mock behavior where possible
• document pricing visibility expectations for role-specific model configurations

Why this matters

If Agent Constitution is going to encourage role-specific model choices, cost needs to be inspectable at the same level as verdicts and audit trails. Otherwise users cannot tell whether a governance pattern is operationally viable.