test(bedrock): use Claude Haiku 4.5 for cache TTL e2e tests

Revert the module-level _CACHE_TTL_MODEL_ID constant from Sonnet 4.5 back
to Haiku 4.5 (us.anthropic.claude-haiku-4-5-20251001-v1:0). Haiku 4.5 is
the newest Haiku available on Bedrock (no 4.6/4.7 yet), officially
documented for 1h TTL support alongside Sonnet 4.5 and Opus 4.5, and runs
the tests noticeably faster with lower per-run CI cost.

The constant is kept so future Haiku versions can be adopted in one line
once they become available.

Author: kpx-dev

tests_integ/models/test_model_bedrock.py

@@ -11,10 +11,12 @@
 
 # Model ID used for prompt-caching TTL integration tests. Per
 # https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html
-# the models that officially support 1h TTL on CachePoint are
-# Claude Opus 4.5, Claude Haiku 4.5, and Claude Sonnet 4.5. Using Sonnet 4.5
-# here for longer expected support lifecycle than Haiku.
-_CACHE_TTL_MODEL_ID = "us.anthropic.claude-sonnet-4-5-20250929-v1:0"
+# the models that officially support 1h TTL on CachePoint are Claude Opus 4.5,
+# Claude Haiku 4.5, and Claude Sonnet 4.5. Haiku 4.5 is the newest Haiku
+# available and is preferred for CI due to lower latency and cost relative to
+# the same-version Sonnet 4.5. Bump this when a newer Haiku is released that
+# supports CachePoint TTL.
+_CACHE_TTL_MODEL_ID = "us.anthropic.claude-haiku-4-5-20251001-v1:0"
 
 
 @pytest.fixture
@@ -596,7 +598,7 @@ def test_prompt_caching_cache_tools_ttl():
     prefix because Bedrock's tool-prefix cache threshold varies by model and region.
     The critical behavior under test here is that the TTL field is accepted end-to-end.
 
-    Uses Claude Sonnet 4.5 which supports TTL in CachePointBlock on Bedrock per
+    Uses Claude Haiku 4.5 which supports TTL in CachePointBlock on Bedrock per
     https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html
     (Claude Opus 4.5, Claude Haiku 4.5, and Claude Sonnet 4.5 all support 1h TTL).
     """
@@ -633,7 +635,7 @@ def test_prompt_caching_cache_config_auto_with_ttl():
     Verifies that the cache point appended to the last user message by _inject_cache_point
     carries the configured TTL, and that Bedrock accepts the request.
 
-    Uses Claude Sonnet 4.5 which supports TTL in CachePointBlock on Bedrock per
+    Uses Claude Haiku 4.5 which supports TTL in CachePointBlock on Bedrock per
     https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html
     """
     model = BedrockModel(
@@ -672,7 +674,7 @@ def test_prompt_caching_aligned_1h_ttl_across_checkpoints():
     This test sets 1h TTL on all three checkpoints simultaneously and verifies the
     call succeeds.
 
-    Uses Claude Sonnet 4.5 which supports 1h TTL per
+    Uses Claude Haiku 4.5 which supports 1h TTL per
     https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html
     """
     model = BedrockModel(