Recognize all Claude model ids in encode_prompt

[wmeddie]

Problem

encode_prompt() only builds the Bedrock Messages payload when the model id starts with anthropic.claude-3 or apac.anthropic.claude-3:

if model_id.startswith('anthropic.claude-3') or model_id.startswith('apac.anthropic.claude-3'):

Claude 4.5 inference-profile ids match neither — different region prefix (jp.) and version (4-5, not 3):

• jp.anthropic.claude-sonnet-4-5-20250929-v1:0
• jp.anthropic.claude-haiku-4-5-20251001-v1:0

So encode_prompt returns an empty messages list and every InvokeModel call fails with messages: at least one message is required. With the comma-separated fallback from this branch, both candidates fail and the agent dies:

[bedrock-agent] all models failed: [
 ('jp.anthropic.claude-sonnet-4-5-20250929-v1:0', "ValidationException(... at least one message is required)"),
 ('jp.anthropic.claude-haiku-4-5-20251001-v1:0', "ValidationException(... at least one message is required)")]

This breaks any downstream agent that adopts the 4.5 model ids.

Fix

Match any anthropic.claude model id instead, covering all region prefixes (jp./apac./us./eu./none) and versions. The Messages content-block format (and the system → user rewrite for mid-conversation messages) is identical across Claude 3+ models.

-    if model_id.startswith('anthropic.claude-3') or model_id.startswith('apac.anthropic.claude-3'):
+    if 'anthropic.claude' in model_id:

Verified encode_prompt now returns non-empty messages for both 4.5 ids and the legacy anthropic.claude-3-5 / apac.anthropic.claude-3-5 ids (regression-safe), while non-Claude ids (e.g. Titan) still return empty.

Based on bedrock-model-fallback (PR #10) since that branch introduced the model ids that expose this.