[plugins] port Amazon Bedrock to Python

[cabljac]

Port of the Amazon Bedrock plugin to Python, using the Go plugin
(genkit-ai/aws-bedrock-go-plugin)
as the source of truth. Same slice discipline as the Go train
(genkit-ai/aws-bedrock-go-plugin#136).

Landing strategy

Work lands as small PRs into a shared base branch, py-bedrock-plugin,
rather than straight to main. The base branch rebases onto main
periodically; when the train completes, one PR takes the assembled plugin
to main. Reviewers get small diffs during development and the full
plugin at the end.

Plugin lives at py/plugins/amazon-bedrock, package
genkit-plugin-amazon-bedrock, module genkit.plugins.amazon_bedrock.

Port structure (Go file → Python module)

Go file	Responsibility	Python module
bedrock.go	plugin init, model/embedder definition, cache-point part	plugin.py
types.go	config, roles, finish reasons, reasoning part ctor	config.py, typing.py
models.go	capability registry, inference-profile prefix stripping	model_info.py
generate.go	Converse request/response conversion	converters.py + models.py
stream.go	ConverseStream block-indexed reassembly	models.py
embed.go	Titan/Cohere/Nova embedders, batching, concurrency	embedders.py
image.go	Titan Image / SD / Nova Canvas via InvokeModel	image.py
rerank.go	Cohere Rerank via InvokeModel	rerank.py

Key decisions

• Namespace bedrock, matching Go, so model refs read
  bedrock/anthropic.claude-... in every language.
• Plain boto3 + thread bridge, behind an internal transport seam.
  aioboto3 rejected: aiobotocore pins botocore to a narrow patch range, so
  the plugin would dictate boto3 versions for every user. A thread bridge
  (asyncio.to_thread for unary calls, a pump thread for ConverseStream)
  keeps the event loop unblocked at LLM concurrency levels. Same call as
  Strands, pydantic-ai (Should we move to aioboto3? pydantic/pydantic-ai#1222), and langchain-aws.
  All boto3 calls go through one transport module so we can swap to AWS's
  async SDK when it matures.
• Reranker as a plugin helper. Neither runtime has a usable reranker
  primitive. Python core carries ActionKind.RERANKER
  (_core/_action.py:128) and nothing else: no define_reranker, no
  ai.rerank(), and the schema types are denylisted from codegen
  (py/scripts/schema_to_typing.py:17-28). Go has no reranker action type
  at all, though it does ship the generated
  ai.RerankerRequest/RerankerResponse types. Registering an action of
  that kind works mechanically, the registry does not constrain kind, but
  callers would be left on resolve_action plus arun with hand-rolled
  types. So rerank ships as a method on the plugin, as in Go. Worth
  revisiting if ai.rerank() lands upstream, since a registered action
  would also pick up reflection listing and tracing spans.

PR train

• 1. Skeleton: scaffold, pyproject, plugin class, config types -feat(py/plugins/amazon-bedrock): scaffold the plugin package #6050
• 2. Model capability registry + inference-profile handling - feat(py/plugins/amazon-bedrock): add model capability registry #5699
• 3. Generate sync path (Converse): converters, tools, tool choice,
  media, cache points, usage, stop reasons - feat(py/plugins/amazon-bedrock): add Converse non-streaming generate path #5876
• 4. Streaming (ConverseStream) + reasoning round-trip - feat(py/plugins/amazon-bedrock): add ConverseStream streaming path #5970
• 5. Embedders (Titan, Cohere, Nova) - feat(py/plugins/amazon-bedrock): add embedders (Titan, Cohere v3, Nova) #6010
• 6. Image generation - feat(py/plugins/amazon-bedrock): add image generation (Nova Canvas, Stability) #6011
• 7. Reranker helper - feat(py/plugins/amazon-bedrock): add rerank helper (Cohere, Amazon) #6046
• 8. Sample app + README - feat(py/plugins/amazon-bedrock): finish the sample app and README #6048
• 9. Docsite page

Tests land with each slice: pytest-asyncio unit tests plus env-gated live
tests mirroring the Go *_live_test.go matrix.

Open questions

• boto3 version floor: Converse cache points and newer stop reasons need
  recent botocore; will pin it down at slice 3.
• Go train items 11/13/14 (region resolution, live-test matrix, docs) are
  still open upstream; we'll port their outcomes when they land rather
  than freezing current Go behavior.

[cabljac]

Two follow-ups on the model registry from slice 2 review (#5699 (comment)), both better as a follow-up on top of the landed stack than an amend to slice 2, since model_info.py is also touched by slices 3, 5 and 6.

• Registry stops at 4.6 (model_info.py:65-66). claude-opus-4-7, claude-opus-4-8, claude-sonnet-5, claude-opus-5 and claude-fable-5 are ACTIVE on Bedrock and missing. They resolve through the unknown-model default, so the only effect is a stage: unstable badge. IDs should come from aws bedrock list-inference-profiles rather than inferred, since 4.6 dropped the date stamp and the -v1 suffix is inconsistent (opus-4-6-v1 has it, sonnet-4-6 does not). Same gap exists in the Go plugin's models.go, which is where the registry was verified against.
• model_info.py:151,171 sets label=model_name. Peers use "<Provider> - <Model>", so the Dev UI picker shows bare IDs next to Anthropic - ... and Ollama - ... entries.

[cabljac]

Corrected the reranker bullet above. The original said Python core has ActionKind.RERANKER so we register a real reranker, but that enum member is the only rerank artifact in core, and #6046 shipped a plugin method instead. Also relabelled slice 7 to match.