Native tool-call SFT + eval, transcript-distillation trajectories, local-serve + Codex skill sync

[ProfSynapse]

What's included

• SFT trainer: config-driven loss_mask_mode + tool_call_mode (assistant_only / native), plus a single-shard load fix that skips split-count verification when loading one shard.
• Transcript distillation: native tool-trajectory emit, with claude.ai export and Codex CLI adapters for mining local agent transcripts into SFT/KTO rows.
• Evaluator: prompt-set support, per-case verifier dispatch, and Qwen XML tool-call response parsing.
• tuner: local-serve vLLM handler (tuner/handlers/local_serve_handler.py) and a Docker utility module (tuner/utils/docker.py).
• Skills: .codex/skills sync mirror added alongside .skills / .agents/skills / .claude/skills, plus .gitignore hygiene and a configs/transcript_import/default.yaml import config.

What's intentionally NOT included

• Tools/ holdout-materialization + aggregator scripts and their tests (tests/tools/, tests/evaluator/) — these are personal-model glue that lived on the feature branch and are out of scope for this generic PR.
• The runner-dispatch test is left out because it imports the personal aggregator.

Testing

New tests added under:

• tests/trainers/sft/ — loss-mask modes, tool-call mode data loader, native tool calls, preprocessing contract, chat-template kwargs passthrough, multi-split data loader.
• tests/scripts/test_trajectory_emit.py — transcript trajectory emit (36 passing).
• tests/shared/ — verifier extraction and verifier dispatch.
• tests/cloud/test_training_capacity.py — training capacity.

🤖 Generated with Claude Code