WIP: PD-DFlash — B2 serving-path §8 measurement + go/no-go

[drunkcoding]

WIP: PD-DFlash — B2 serving-path §8 measurement + go/no-go

DRAFT — do not merge. Closes out the Phase-A §8 go/no-go by measuring B2
(the co-designed 2-D verify-scheduler baseline) and a real kv_occupancy_bytes
through the continuous-batching serving path, replacing the
BLOCKED_UNTIL_2D_SCHEDULER placeholder. Depends on #164 (native stat
accessors) and Task 6 (SpecSession + 2-D verify scheduler + engine wiring),
both already on dev.

What this PR does (runner wiring only — no engine/scheduler/session change)

Task 6's engine wiring is already merged, so B2 needed only benchmark-runner
wiring — no architectural change:

• build_b2_serving_config() (pd_dflash_serving.py): sets all four Task-6
  verify budgets so Scheduler.verify_scheduling_enabled flips
  ContinuousBatchingEngine onto _step_speculative_session (the 2-D admission
  scheduler). Token budget/cap = block_size; expert-byte budget/cap are
  unbounded — exactly B2's "token-deficit scheduler and no expert-byte
  coupling" (BASELINES['B2']).
• measure_configuration_serving() (_serving_measure.py): drives the
  engine via add_request/run_until_done (NOT MoE.generate()), captures
  real acceptance from each VerifyResult, resets native expert-cache counters
  around the timed window, and reads a real kv_occupancy_bytes by
  peak-sampling the serving PagedKVCache allocator. (The sync generate()
  path has no serving KV manager, hence B0/B1/B3 honestly report 0.0.)
• run_experiment routes --baseline B2 through the serving path via
  SERVING_BASELINES; everything else keeps the existing MoE.generate() path.
• CPU test asserts the B2 config enables the token-only verify scheduler.

Opt-in and GPU-gated; default serving/hot-path behaviour is unchanged.

Measurement — gpt-oss-20b, B=16, concurrency=1, one RTX PRO 6000 (SM120), measured H2D 57.6 GB/s

Primed FP4 offload dir reused; B2 measured via serving path this PR; B0/B1/B3
are the prior real rows (they are defined on the sync path and keep it).

baseline	tok/s	goodput@SLO	accept a	TTFT s	hit rate	route-ahead cov	wasted B	expert occ (GB)	kv occ (B)
B0 (AR, offloaded)	18.51	18.51	1.00	0.084	0.978	0.0	0	19.97	0 (no serving KV)
B1 (DFlash, AR prefetch)	62.07	62.07	1.00	0.081	0.894	1.0	0	18.86	0 (no serving KV)
B2 (2-D verify sched.)	79.81	79.81	7.98	0.123	0.999	1.0	0	20.55	1 572 864 (real)
B3 (resident upper bound)	84.56	84.56	1.00	0.084	0.890	1.0	0	18.45	0 (no serving KV)

Go/No-Go verdict (computed with the repo's own report.py)

• report.validate_result_matrix — PASS (all B0–B3 present, every §8 metric finite/non-negative; no blocked baseline).
• report.evaluate_matrix — ok, blocked=[] missing=[] invalid=[].
• Money-plot ordering B0 < B1 < co-design(B2) → B3: HOLDS — 18.51 < 62.07 < 79.81 < 84.56.
• Route-ahead beats AR (B0→B1): yes (+235%).
• Co-design B2 beats B1: yes (+28.6%), and closes 66% of the B1→B3 (resident) gap while remaining fully offloaded.
• B2 kv_occupancy_bytes real (serving KV manager present): yes (1.57 MB, no fell back to 0.0 warning).

VERDICT: GO. The matrix is complete and all-real; the design §8 ordering holds; the co-design B2 delivers a measurable win over B1 on the serving path.

Scope notes / honest limitations

• Hide inequality (§7): not evaluable from the Phase-A §8 schema — it needs
  t_draft / t_router (BM1, Task 7) and overlap (BM4, Task 10), which are
  not part of the B0–B3 row schema. Reported as out-of-scope for this B2 slice,
  not as a pass/fail.
• Task 9 priority band (ship_priority_band) is a BM3-gated C++ hop and is
  not decided here (no C++ was touched). The strong route_ahead_coverage=1.0
  • hidden-fetch evidence (B2 approaches the resident B3) justifies proceeding
    to the BM3 three-way ablation, but the ship decision remains BM3's
    (bm3_decision), per "no C++ change ships without its BM".
• B0/B1/B3 acceptance shows 1.0 because the sync path did not surface
  step_trace; B2's 7.98 is the real per-round acceptance from VerifyResult.

Verification

• ruff + ruff-format clean; LSP clean on changed files.
• CPU regressions: test_pd_dflash_serving_contract, test_pd_dflash_report,
  serving test_sequence/test_batch/test_scheduler/test_dflash_deficit_scheduler,
  dflash test_engine_wire/test_route_ahead_wire/test_spec_session — all pass.
• gpt-oss offload no-regression: tests/test_gpt_oss_offload_topology.py — pass.
• B2 GPU run reproduced twice (smoke + 64-request), both real, no fallbacks.