Purpose of the change

Fix two issues that prevent MemMachine evaluation benchmarks from running out of the box:

1. the Bedrock reranker requires explicit AWS credentials in the config YAML, blocking SSO/profile/login-based auth
2. the eval helper uses the old LongTermMemoryParams class which was replaced by a backend-discriminated union in Wire EventMemory as backend for LongTermMemory #1395.

Description

Bedrock credential chain fix (reranker_conf.py + reranker_manager.py):

AmazonBedrockRerankerConf previously defined its own aws_access_key_id and aws_secret_access_key fields as required (Field(...)), meaning any config using the Bedrock reranker had to include explicit credentials — SSO, AWS CLI login, and profile-based auth couldn't work. This PR replaces the hand-rolled fields with AWSCredentialsMixin (which defaults to None with env-var fallback) and updates RerankerManager._build_amazon_bedrock_reranker to only pass credentials to boto3.client() when they're actually set. When omitted, boto3 falls through to its default credential chain (~/.aws/credentials, ~/.aws/config, SSO, instance metadata).

Eval script fix (agent_utils.py):

PR #1395 ("Wire EventMemory as backend for LongTermMemory") changed LongTermMemoryParams from a single concrete class to a Discriminator-based union of DeclarativeBackendParams | EventBackendParams. The evaluation helper init_memmachine_params still used the old LongTermMemoryParams, which no longer accepts the fields directly. This PR updates it to use DeclarativeBackendParams(backend="declarative", ...).

Tests (test_reranker_conf.py):

• Added test_valid_amazon_bedrock_reranker_conf_without_explicit_credentials — verifies the config accepts no AWS credentials when env vars are unset
• Added test_valid_amazon_bedrock_reranker_conf_uses_env_credentials — verifies env-var resolution populates credentials correctly
• Removed test_missing_required_field_in_bedrock_reranker — this tested the old required-field behavior which no longer applies

Fixes/Closes

None

Type of change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Refactor (does not change functionality, e.g., code style improvements, linting)
• Documentation update
• Project Maintenance (updates to build scripts, CI, etc., that do not affect the main project)
• Security (improves security without changing functionality)

How Has This Been Tested?

• Unit Test
• Test Script

Lint (all changed files):

uv run ruff check evaluation/utils/agent_utils.py \
  packages/server/src/memmachine_server/common/configuration/reranker_conf.py \
  packages/server/server_tests/memmachine_server/common/configuration/test_reranker_conf.py \
  packages/server/server_tests/memmachine_server/common/resource_manager/test_reranker_manager.py

Unit tests (config + manager):

uv run pytest \
  packages/server/server_tests/memmachine_server/common/configuration/test_reranker_conf.py \
  packages/server/server_tests/memmachine_server/common/resource_manager/test_reranker_manager.py

End-to-end eval (LoCoMo benchmark, retrieval_agent mode):

Tested the "no keys in config" path using a personal AWS account via boto3's default credential chain. The Bedrock reranker (amazon.rerank-v1:0, us-west-2) resolved credentials automatically and completed successfully — no AccessDeniedException, no ValidationError. All 10 conversation groups processed with the retrieval agent routing queries through MemMachineAgent, SplitQueryAgent, and ChainOfQueryAgent as expected.

./run_test.sh locomo exp1 search retrieval_agent