fix(cost): reserve reasoning against an explicit total completion cap

[drewstone]

Problem

Paid-call admission currently receives ChatRequest.maxTokens and can treat it as the maximum billable completion. That assumption is false for providers where max_tokens limits visible output while hidden reasoning is billed separately.

Direct provider proof on 2026-08-10, without a coding harness:

• model: glm-5.2
• request: max_tokens: 8
• response: 3 visible tokens, 135 completion tokens, 132 reasoning tokens
• request with max_completion_tokens: 256: 119 completion tokens, including 116 reasoning

This is distinct from CLI harness prompt overhead tracked in #544.

Completion criteria

• Give cost reservation an explicit maximum total billed completion-token value.
• Keep any visible-output and reasoning limits separate in request/evidence types.
• Include the total cap in maximum-charge calculation before dispatch.
• Reject capped paid execution when the caller/runtime cannot prove an enforceable total cap.
• Compare reported completion plus reasoning semantics against the admitted limit without double counting.
• Preserve separate visible, reasoning, cached, and total token evidence.
• Keep execution transport in agent-runtime as required by refactor(execution): remove Eval-owned paid model transports #539.
• Add a regression where visible output is below maxTokens but billed reasoning exceeds it.