Release v1.4: Grid188 decode and runtime synchronization fixes

[davidsyoung]

Summary

• Build v1.4 directly on the immutable public v1.3 image digest.
• Default-enable the exact M4 heterogeneous W4A16 Grid188 path, with direct128 and serial fail-closed fallbacks.
• Carry the fixed FC1/FC2 tile geometry through the registered custom-op boundary.
• Remove recurring MTP/DCP metadata synchronizations and redundant MXFP8 scale initialization.
• Release warmup-only allocator cache once, retain AOT mode without loading/saving the serialized standalone-AOT function, and reduce avoidable Triton specialization/temp storage.
• Replace the experiment history in the docs with the current mechanisms and verified benchmark results.

How the optimizations work

Area	Change
Hybrid MoE decode	One 188-CTA grid executes NVFP4+NF3 FC1, gated SiLU, and weighted FC2 accumulation from global expert IDs. Exact device/resource/tensor admission protects the path.
MTP metadata	A CPU-resident optimistic sequence-length bound is propagated through draft-step attention metadata, avoiding a lazy device-to-host length copy.
DCP metadata	The constant rank-offset tensor is cached by device/rank instead of rebuilt on CUDA every call.
Dense MXFP8	Scale buffers skip unity initialization immediately before the quantizer overwrites every logical scale used by GEMM.
Runtime lifecycle	Unused allocator-cache blocks are released after warmup/capture; AOT stays enabled while standalone serialized-function load/save is bypassed.
Auxiliary kernels	Route count remains a runtime Triton argument, and native small-batch filtering uses bounded Triton workspace when available.

The hybrid decode and post-warmup paths are enabled in source and in the image. Explicit 0/false environment overrides remain available for diagnosis; normal operation no longer depends on Compose-only gates.

Benchmarks

Measured on 4× RTX PRO 6000 Blackwell 96 GB, TP4+DCP4, MTP3 probabilistic draft sampling:

• Exact-32k prefill: 2,128 / 2,128 tok/s, versus the public v1.3 reference of 2,131.
• Deterministic C1/ctx0, five 60 s cells: median 105.703 client aggregate / 112.153 per-user active-decode tok/s.
• Stochastic C1/ctx0: 102.895 aggregate / 107.575 per-user tok/s.
• Stochastic C8 aggregate at 0 / 32k / 64k: 318.535 / 307.744 / 293.774 tok/s.
• KV pool: 637,440 tokens.

Aggregate and per-user throughput are distinct metrics. Deterministic and stochastic payloads are reported separately, not compared as an A/B. BENCHMARKS.md contains the run-level values and interpretation limits.

Verification

• Built the final Dockerfile successfully from the pinned public v1.3 digest; all nine overlays passed in-image py_compile.
• docker compose config -q passed for the final Compose file.
• Standalone image environment confirmed the three hybrid gates, post-warmup cache release, and the two AOT lifecycle defaults are image-owned and enabled.
• Repository-wide publication scan found no private/internal naming, hosts, image tags, or artifact paths.

No new regression-test framework is included; the source-equivalent runtime stack is covered by the recorded full-model benchmark evidence, and the PR packaging was verified through the final image build.