Enhance SQS monitoring and timeout settings with documentation updates

[gafnts]

This pull request improves the reliability and accuracy of load test reporting and cleanup for the agentic flavor deployment. The main changes address issues with test harness timeouts, cleanup of incomplete documents, and the clarity of concurrency cap reporting. These updates help ensure that test results more accurately reflect the system's real behavior and reduce the risk of spurious failures due to measurement artifacts.

Load test harness and reporting improvements:

• Increased the default timeout in the await_completion function from 600s to 900s in harness.py, ensuring the harness waits long enough for documents requiring SQS retries to reach a terminal state, thus reducing false timeouts that can fail SLO 1.
• Enhanced the docstring for await_completion to clarify the relationship between the timeout, SQS visibility, and SLO correctness.
• Updated the cleanup function in harness.py to skip deletion of documents still marked as timeout, preventing premature deletion that could cause phantom DLQ failures due to in-flight retries encountering missing S3 objects. Added a printout to inform the user of skipped documents.

SLO and concurrency reporting updates:

• Improved the concurrency cap SLO logic in report.py to clarify that Lambda concurrency and throttles are the authoritative signals, while SQS in-flight is a noisy proxy now reported for reference but not used for gating. Updated SLO output to include SQS in-flight peak as a proxy value. [1] [2]

Documentation and findings:

• Expanded the ADR post-implementation section with detailed results and findings from burst and sustained load runs, documenting SLO verdicts, root causes of failures, and rationale for recent harness and measurement changes.