[DataFlow runtime] Phase B3 — domain Trainer wrapping the runtime seam

[maocheng23]

Phase B (domain abstractions) — 3/3. Stacked on #632 (B2).

Introduces the domain training layer specforge/training/ with a caller-facing Trainer that composes the whole spine behind one object + .fit():

FeatureDataLoader  +  FSDPTrainingBackend.prepare_model (FSDP wrap)
                   +  spec.make_strategy -> TrainerCore -> TrainerController

Trainer is the canonical assembler now; launch._assemble_trainer delegates to it and returns the same (TrainerController, FeatureDataLoader) tuple, so every build_*_runtime path is byte-for-byte unchanged (one wiring path, no fork). The runtime seam (TrainerController / TrainerCore / DraftTrainStrategy / FSDPTrainingBackend) is untouched — this is the domain facade over it. Topology (offline/online/disagg) stays invisible to Trainer; it's absorbed by the (ref source + FeatureStore) it's handed. No HiddenStateStream — the loader is the stream.

• specforge/training/{__init__.py, trainer.py} (new)
• launch.py: _assemble_trainer delegates to Trainer; drops the now-unused FeatureDataLoader/FSDPTrainingBackend/ParallelConfig/TrainerCore/TrainerController imports.
• New test: tests/test_runtime/test_domain_trainer.py — fakes the runtime pieces and asserts the composition (refs enqueued, loader/backend/core/controller args, ack_fn wired to the DataFlowController, .fit() delegates over the loader).

Validation

Full tests/test_runtime 214 OK (2 skip, 1 xfail) on 8×H200 — the existing launch/equivalence tests now run through this Trainer, so byte-identical loss is covered. Adversarial review: 0 confirmed defects.

Known gap (not blocking)

No single test drives the online loop with a real SGLang target (backend="sglang" → capture backend → adapter → RolloutWorker → fit); the sglang capture is validated end-to-end at the capture level (parity test) and the online rollout→train loop is validated with an HF target, but not combined. Can add on request.

🤖 Generated with Claude Code

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!