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 removes the client_generator technical debt. The tinker client is untouched — only documented for its Gateway-based, generation-only multi-turn path.

Changes

Technical debt cleanup

• Remove client_tools/client_generator.py and convert the affected client modules (dataloader / dataset / processor / model / sampler) to hand-maintained sources by stripping their AUTO-GENERATED headers (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 keeping 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_bridge. vLLMSampler now JSON-safes pre-encoded InputFeature inputs before the POST.

Embedding training

• processor/base.py: rebuild a position_ids reference for last-valid-token pooling when HF omits position_ids in a standard padded batch.
• megatron.py set_loss: inject the DP group into loss.process_group so the InfoNCE in-batch-negative all-gather stays within DP ranks under PP (prevents an NCCL deadlock on earlier pipeline stages that never enter the loss).

MockSampler multi-turn knobs

• Add opt-in new_input_feature plus configurable stop_reason / tool_call_text / tool_call_turns so a local CPU-only multi-turn e2e can run without a GPU. Backward compatible by default.

Tests

• Embedding: CPU-only mock protocol/link validation; GPU-gated real-model loss finiteness, embedding-patch auto-rollback, bare-library parity, pos_sim upward trend, and SP/CP consistency.
• Multi-turn: CPU-only property/unit tests, a local mock-sampler HTTP e2e, and GPU-gated behavioural-alignment + Gateway no-logprobs tests.

Docs & cookbook

• Client embedding training example and client multi-turn rollout (GRPO) example, both verified against a live megatron server (pp=2/dp=2).
• Document the client embedding call sequence and the Tinker Gateway indirect multi-turn approach with its capability boundaries.

Verification

• Local CPU suite passes; GPU-gated cases skip cleanly on machines without a GPU.

[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.