WIP: PD-DFlash — native offloaded-expert stat accessors (unblock §8 B0–B2)

[drunkcoding]

WIP / DRAFT — do not merge

Wires the native offloaded-expert statistics accessors so the PD-DFlash Phase A §8 matrix's B0–B2 (offloaded) rows emit REAL metrics instead of 0.0 fallbacks — the last blocker to computing the hardware go/no-go. All accessors are read-only, additive, opt-in; no default/hot-path behavior changes.

Root cause of the prior blocker (measured on origin/dev)

The gpt-oss offloaded baselines were unmeasurable because _count_offloaded_experts(engine) returned 0 (gpt-oss's packed experts leave expert_prefetcher.expert_nbytes_map empty), so require_offloaded(B0|B1|B2, 0) raised. gpt-oss experts flow through the native ExpertDispatcher (FP4), whose hit/occupancy accounting was never surfaced to Python — hence expert_cache_hit_rate/expert_occupancy_bytes/wasted_prefetch_bytes all fell back to 0.0.

Counters exposed (C++ → pybind → Python)

Metric	Native source (read-only)	pybind	Python surface
expert_cache_hit_rate	ExpertDispatcher passive hit/access counters (gpt-oss); NodeBody hit/visit aggregate (others)	expert_dispatcher.get_cache_hit_rate / prefetch_handle.get_hit_rate	ExpertPrefetcher.get_hit_rate(), engine.expert_cache_hit_rate
expert_occupancy_bytes	Σ byte_size over cached_experts_ (dispatcher) + on-CUDA node byte_size (handle)	get_cache_occupancy_bytes / get_expert_occupancy_bytes	ExpertPrefetcher.expert_occupancy_bytes(), engine.expert_occupancy_bytes
wasted_prefetch_bytes	Σ byte_size × unused_count (topology)	get_wasted_prefetch_bytes	ExpertPrefetcher.wasted_prefetch_bytes(); runner native fallback
num_offloaded_experts	is_tensor_offloaded over expert_tensor_map (populated for gpt-oss)	(existing is_tensor_offloaded)	ExpertPrefetcher.num_offloaded_experts, engine.num_offloaded_experts
measured_h2d_bytes_per_second	opt-in on-device H2D probe	—	--probe-h2d flag (flag --measured-h2d-gbps already accepted)
kv_occupancy_bytes	engine KV-cache manager when present	—	engine.kv_occupancy_bytes (see note)

The dispatcher hit/access counters are two std::atomic increments at the existing exec-arg emit sites (same instrumentation class as gpu_hit_cnt++/incache_visit_count++), reset by ClearExpertCacheCounts(); they change no routing/dispatch behavior.

Qwen 404 fix

Qwen/Qwen3-Coder-30B-A3B (harness default) returns HTTP 404 / never existed (verified via HF API). Replaced with the real, ungated Qwen/Qwen3-Coder-30B-A3B-Instruct (qwen3_moe, per-expert tensors) in the frozen §8 contract + tests + docstring. Draft z-lab/Qwen3-Coder-30B-A3B-DFlash unchanged (verified present).

RE-RUN §8 slice (SM120, gpt-oss-20b, --device-memory-ratio 0.30, --probe-h2d)

report.py --allow-blocked B2 → exit 0. B0/B1 now RUN (no require_offloaded RuntimeError); B2 remains the runner's pre-existing BLOCKED_UNTIL_2D_SCHEDULER.

baseline	expert_cache_hit_rate	expert_occupancy_bytes	wasted_prefetch_bytes	measured_h2d B/s	tok/s
B0	0.978	19.97 GB	0.0 (real)	57.6 GB/s	18.51
B1	0.894	18.86 GB	0.0 (real)	57.6 GB/s	62.07
B3	0.890	18.45 GB	0.0	57.6 GB/s	84.56

Before: B0–B2 could not run at all; B3 alone reported hit_rate/occupancy/kv = 0.0 fallbacks. 4 of the 5 named gaps now emit real values.

Note on kv_occupancy_bytes: still 0.0 in this run — honestly, not fabricated. The MoE.generate() path uses HF's own KV cache (no MoE-Infinity KV manager), so the engine has nothing resident to report. The accessor is wired to report real KV occupancy when a KV-cache manager is present (continuous-batching/serving path).

Verification

• New CPU tests tests/python/dflash/test_native_stat_accessors.py (14) — accessor surface with _store fakes.
• Full dflash CPU suite: 371 passed, 7 GPU-gated skips.
• No-regression tests/test_gpt_oss_offload_topology.py (+policy): 10 passed.
• ruff + LSP clean on all changed files. _store rebuilt for sm_120, new symbols import cleanly.

Guardrails

Read-only accessors only; no GLM / feat/dflash-spec-decode / #119 touched; not on dev/main; DRAFT — do not merge.