fix(kotlin): Honor Retry-After for pool warmups

[Gujiassh]

Summary

• Preserve SandboxRateLimitException during Kotlin pool warmup completion and pause new warmups for Retry-After, with a 10-second default and the existing 60-second SDK cap.
• Keep rate-limit throttling separate from degraded failure accounting, extend concurrent deadlines monotonically, and schedule a reconcile as soon as the throttle expires.
• Keep idle reaping and shrink active during throttling, reset the local throttle on pool restart, and document the behavior.

The rate-limit state is scoped to the active pool run, avoiding a shared PoolStateStore contract change and keeping this fix separate from the windowed degraded-detection work in #1499.

Fixes #1500.

Testing

• Not run (explain why)
• Unit tests: ./gradlew :sandbox:test (324 tests, 0 failures)
• Integration tests
• e2e / manual verification: focused one-second Retry-After recovery and stop/start reset tests
• ./gradlew spotlessCheck
• pnpm docs:build

Breaking Changes

• None
• Yes (describe impact and migration path)

Checklist

• Linked Issue or clearly described motivation
• Added/updated docs (if needed)
• Added/updated tests (if needed)
• Security impact considered
• Backward compatibility considered

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a499814066

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Keep rate-limit pool behavior aligned across SDKs

This branch makes HTTP 429 warmup failures stop contributing to degraded backoff only in Kotlin, but the same client-pool warmup exists in Python and Go and their reconcilers still count every thrown warmup error into record_failures/recordFailures (checked sdks/sandbox/python/src/opensandbox/_pool_reconciler.py:161-164 and sdks/sandbox/go/pool_reconciler.go:271-282). Without the same handling or a documented platform constraint, server 429/Retry-After produces different quota/back-pressure semantics across SDKs.

AGENTS.md reference: sdks/AGENTS.md:L122-L122

Useful? React with 👍 / 👎.