[DataFlow runtime] Phase B4 — adopt the de-EAGLE3 surface (cutover + docs + gate)

[maocheng23]

Phase B — 4/4 (clean-up). Stacked on #633 (B3). Makes the Phase B abstractions load-bearing instead of forward-only scaffolding: after B1–B3 the new generic surface (get_target_engine, TargetEngine.capture) existed but no production caller used it yet — callers still hit generate_eagle3_data / get_eagle3_target_model. This adopts it.

1. adapter cutover — SGLangAdapter / DFlashAdapter now call the engine's generic capture(...) instead of generate_eagle3_data / generate_dflash_data (kept as byte-identical back-compat aliases). This is the dataflow runtime's real capture path, so TargetEngine.capture() is now exercised end-to-end.
2. factory cutover — the dataflow test fixtures + online launch tests build targets via get_target_engine(strategy=, backend=); get_eagle3_target_model / get_dflash_target_model stay as shims (legacy scripts untouched).
3. docs — runtime/inference/DESIGN.md, runtime/ARCHITECTURE.md, and the adapter docstrings now describe the boundary as TargetEngine.capture (sglang glue behind SGLangCaptureBackend) instead of Eagle3TargetModel.generate_eagle3_data.
4. explicit B gate — tests/test_runtime/test_phase_b_gate.py: capture() ≡ legacy generate_*_data() bytewise (eagle3 + dflash), a stable captured-feature digest, and per-step-loss reproducibility through the full online capture→train path — instead of relying only on the full-suite's indirect coverage.

No behavior change: capture() dispatches to the legacy method.

Validation (8×H200)

Full tests/test_runtime 218 OK (2 skip, 1 xfail) = the prior 214 + the 4 new gate tests. The new gate confirms capture() is byte-identical to the legacy extraction for both eagle3 and dflash.

With this, Phase B is complete. (Deferred to Phase E by decision: collapsing the per-(algorithm × backend) engine classes into per-backend generic engines + a CaptureSpec, alongside the models/targets/ move — see plan.md §2.3.)

🤖 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!