fix: update timeout of RLVR sequence_length test

[rsareddy0329]

Issue #, if available:

Description of changes:

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[github-actions]

Claude finished @rsareddy0329's task in 24s —— View job

---

PR Review

• Read PR diff (/tmp/pr.diff)
• Review the change for correctness & best practices

---

The change bumps max_wait_time from 3600 to 7200 seconds in test_rlvr_trainer_lora_with_sequence_length, with a clear explanatory comment noting that sequence_length training takes >1 hour.

Looks good. This is a straightforward, correct fix for a test that was timing out. The polling loop (poll_interval = 30) and terminal-status handling are unchanged and remain correct — the loop will still break early on Completed/Failed/Stopped, so the longer timeout only affects the worst-case wait, not the happy path.

Minor (optional, non-blocking) observations — not required for this PR:

• If the job never reaches a terminal state, the final assert training_job.training_job_status == "Completed" will now surface after 2 hours rather than 1. That's the intended tradeoff here, so no action needed.
• Consider whether an integ test blocking up to 2 hours fits within your canary/CI budget; if this becomes a pattern, a shared timeout constant might be cleaner. Not something to address in this PR.

No correctness, security, or backward-compatibility concerns. 👍

[github-actions]

Claude finished @rsareddy0329's task in 18s —— View job

---

PR Review

• Read PR diff (/tmp/pr.diff)
• Review changes for correctness & best practices

---

This PR makes two small, test-only changes:

1. test_nova_model_customization_deployment.py — adds @pytest.mark.xdist_group("bedrock_deployment") to TestNovaBedrockDeployment. This pins all tests in the class to the same pytest-xdist worker, serializing Bedrock deployment tests so they don't run concurrently and collide on shared Bedrock resources. Sensible fix for flaky parallel runs.

2. test_rlvr_trainer_integration.py — bumps max_wait_time from 3600 → 7200 in test_rlvr_trainer_lora_with_sequence_length, with a clear comment explaining that sequence_length training takes >1 hour. The polling loop (poll_interval = 30) and terminal-status handling are unchanged, so the longer timeout only affects the worst case — the loop still breaks early on Completed/Failed/Stopped.

Looks good. Both are correct, low-risk test infrastructure fixes with no impact on public SDK behavior. No correctness, security, or backward-compatibility concerns. 👍

Minor (optional, non-blocking):

• An integ test that can now block up to 2 hours is worth keeping an eye on for canary/CI budget. If timeouts like this become a recurring pattern, a shared timeout constant might be cleaner — not needed for this PR.