Fix: fail fast on rolling provider usage limits #112

Context

A production code-review run with CLI 0.4.3 and zai-coding-plan/glm-5.3 reached the provider’s five-hour rolling usage limit. The CLI treated that response as retryable, made ten retry attempts, and remained alive until the executor's two-hour deadline.

Problem / Goal

A provider response that says its usage limit will reset hours later cannot be recovered within a normal task-execution deadline. Retrying it keeps executions appearing to run while consuming capacity, then reports a timeout instead of the actionable provider limit.

Reproduction

1. Configure a provider response that is marked retryable and says Usage limit reached for 5 hour with a reset time beyond the task deadline.
2. Run aictrl run --format json.
3. Observe the current session retry loop accepts the long delay and retries until the outer executor timeout.
4. Expected: classify this rolling usage-limit response as terminal, emit the normal structured failure lifecycle, and exit non-zero without sleeping/retrying.

Environment

• CLI: @aictrl/cli 0.4.3
• Provider/model: Z.ai GLM-5.3 Coding Plan
• Downstream executor deadline: two hours

Proposed Approach

In src/session/retry.ts, recognize provider usage-limit/reset responses as non-retryable before the retry delay is accepted. Preserve ordinary transient rate-limit retries, and keep the terminal session_error / invocation completion events and error text available to headless consumers.

Acceptance Criteria

• A response containing Usage limit reached and a future reset is not retried.
• Headless JSON mode emits a structured terminal failure and exits non-zero for that response.
• Normal short-lived 429/rate-limit and overload retry behavior remains covered by regression tests.
• The error makes clear that the provider usage limit, rather than the executor deadline, caused the failure.

Out of Scope

• Changing retry policy for all provider errors.
• Product dashboard/heartbeat instrumentation (handled downstream).
• Retrying automatically after the provider reset time.

Roadmap Alignment

• Pillar: EXEC
• Priority: P1 — production reliability and operator diagnosis.
• Milestone: Enterprise Observability

References

• Related telemetry contract: Expose execution outcome and retry measurement inputs #110 (intentionally excludes retry-policy changes).
• Downstream handoff: aictrl-dev/aictrl#1509.