feat: client-side embedding training & multi-turn rollout support

[Yunnglin]

Summary

Exposes the core-library embedding training (InfoNCE) and trainable multi-turn rollout capabilities through twinkle_client, and cleans up the client_generator technical debt. tinker client stays untouched — only documented for its Gateway-based indirect (generation-only) multi-turn path.

Technical debt cleanup

• Remove client_tools/client_generator.py; convert affected client modules (dataloader/dataset/processor/model/sampler) to hand-maintained sources by stripping their AUTO-GENERATED headers (byte-for-byte otherwise unchanged).

Shared bridge function

• Extract bridge-token stitching into the shared pure function twinkle_agentic/rollout/bridge.py::extend_with_bridge. MultiTurnRollout now calls it instead of maintaining its own copy, so the core-lib and client paths cannot drift.

Client multi-turn rollout

• Add ClientMultiTurnRollout (src/twinkle_client/rollout/) driving HTTP sampling via vLLMSampler.sample(), reusing ToolManager and `extend_with_brid

Technical debt cleanup

• Remove client_tools/client_generator.py; convert affected client modules (dataloader/dataset/processor/model/sampler) to hand-maintained sources be. Backward compatible by default.

Verification

• Embedding: e2e scaffolding + local mock protocol-link validation (Property 1); GPU-gated r

Shared bridge function

• Extract bridge-token stitching into the shared pure function twinkle_agentic/rollout/bridge.py::extend_with_bridge. MultiTurnRollout now calls it instead of maintaining its own copy, so the core-lib and cU-g- Extract bridge-token stme

Client multi-turn rollout

• Add ClientMultiTurnRollout (src/twinkle_client/rollout/) driving HTTP sampling via vLLMSampler.sample(), reusing ToolManager and `extend_with_brid

Technical debt cleanup

• Remove client_totib- Add ClientMultiTurnRollou
  -### Technical debt cleanup
• Remove client_tools/client_generator.py; convert affected client modules (dataloader/dataset/processor/model/samplega- Remove client_tools/clGP

Verification

• Embedding: e2e scaffolding + local mock protocol-link validation (Property 1); GPU-gated r

Shared bridge function

• Extract bridge-token stitching into the shared pure fser- Embedding: e2nn### Shared bridge function
• EOF

'

Notes

• GPU-gated e2e cases require a running server (tests/server/start_e2e_server.py) and cannot be validated locally.

[gemini-code-assist]

Code Review

This pull request introduces client-side multi-turn agentic rollout orchestration (ClientMultiTurnRollout) and adds documentation and tests for embedding training and multi-turn tool calling. Feedback focuses on ensuring proper JSON serialization of PyTorch tensors and template outputs in the HTTP client path, as well as removing an unused variable in the mock sampler.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

The _to_plain function recursively converts numpy arrays to lists, but it does not handle PyTorch Tensor objects. In multimodal scenarios, mm_token_type_ids is padded and concatenated as a torch.Tensor (lines 139-147), which remains a Tensor after _to_plain is called. When the client attempts to serialize this to JSON to send over HTTP, it will raise a TypeError. We should update _to_plain to also convert PyTorch tensors to lists if they are present, using a safe string-based type check to avoid importing torch in environments where it is not installed.

Suggested change

	if isinstance(obj, np.ndarray):
	return obj.tolist()
	if isinstance(obj, np.ndarray):
	return obj.tolist()
	if type(obj).__name__ == 'Tensor' and hasattr(obj, 'tolist'):
	return obj.tolist()

[gemini-code-assist]

Import _to_plain from twinkle_agentic.rollout.bridge so that we can convert the initial pif returned by self.template.encode into a plain dictionary with serializable types before sending it over HTTP.

Suggested change

	from twinkle_agentic.rollout.bridge import extend_with_bridge
	from twinkle_agentic.rollout.bridge import extend_with_bridge, _to_plain

[gemini-code-assist]

The initial pif returned by self.template.encode is not run through _to_plain before being used. If the template returns an InputFeature object containing numpy arrays or PyTorch tensors (which is standard for real templates), calling .setdefault might fail if it's not a dict, and sending it over HTTP on line 139 will cause a JSON serialization error. We should convert the initial pif to a plain dictionary using _to_plain just like the core-library MultiTurnRollout does.

Suggested change

	pif = self.template.encode(traj, add_generation_prompt=True)
	pif.setdefault('messages', list(traj.get('messages', [])))
	pif = self.template.encode(traj, add_generation_prompt=True)
	pif = _to_plain(pif)
	pif.setdefault('messages', list(traj.get('messages', [])))

[gemini-code-assist]

The _MULTI_TURN_KNOBS tuple is defined here but never used anywhere in the codebase. Furthermore, the comment mentions it is 'Kept in a set', but it is defined as a tuple. Since the constructor explicitly defines stop_reason, tool_call_text, and tool_call_turns as keyword-only arguments, Python's native argument binding handles them automatically, making this variable completely redundant. We should remove this unused variable and its comment.