Canonical Fathomless Firmament merge plan and release provenance cleanup

[voipmonitor]

Objective

Make dev/fathomless-firmament and B12X master the complete canonical source
of the next release image. No runtime source change may exist only in a Docker
overlay, release branch, copied site-packages tree, or unpublished local commit.

Why this is required

The v17 image was built from a release integration branch rather than directly
from dev/fathomless-firmament. That branch contained useful direct commits
which had no pull request. For example, fixed-width MTP host synchronization was
removed by commit 83925ac7b6 in the release tree but was absent from canonical
FF. Replacing the image's Python tree with current FF therefore reintroduced a
4.5-5% MTP regression while MTP-off remained unchanged.

The same release history also mixed DSpark correctness fixes, optional
experiments, deployment launchers, memory tuning, and unrelated PCIe work. The
changes below have now been re-extracted onto current canonical branches with
explicit scope and test provenance.

B12X merge order

Merge these into B12X master before building the final image. They touch
independent subsystems, but this order moves runtime correctness prerequisites
ahead of optional formats:

1. B12X PR #28 - handle
   CuTe runtimes where cute.compile is a plain callable. Required for the
   W4A8 dynamic MoE path used by A8 configurations.
2. B12X PR #32 - launch
   the W4A8 resident-grid kernel cooperatively. Prevents graph-capture crashes
   when auxiliary-stream shared-expert work is active.
3. B12X PR #34 - preserve
   preplanned W4A16 tile geometry through the custom-op boundary. Required for
   reliable NF3 packing and launch selection.
4. B12X PR #35 - clean
   current-master NVFP4 MLA KV-cache implementation, including outer scales and
   optional FP8-RoPE records.

NF3 format support from B12X PR #31 is already merged. Draft B12X PR #33 is
superseded by #35 and must not be merged in addition to it.

vLLM merge order

All PRs below are non-draft and target dev/fathomless-firmament directly.

1. #108 - remove the
   duplicate sparse-indexer RoPE quant helper and refresh stale fixtures. Merge
   first because the consolidated DSpark capacity path touches that module.
2. #97 - canonicalize
   the fixed-width MTP host-sync optimization that was previously release-only.
3. #103 - return
   variable-width draft IDs through the existing asynchronous output transfer.
   Merge after #97.
4. #98 - isolate the
   probabilistic draft Gumbel stream from unrelated CUDA RNG consumption.
5. #99 - do not allocate
   speculative lookahead slots to running chunked-prefill steps.
6. #100 - preserve MLA
   KV quantization metadata while normalizing cache page sizes.
7. #101 - harden DS4
   sparse verifier metadata for DCP1 and padded graph replays. This is a DS4
   target fix and is intentionally separate from DSpark.
8. #102 - give DS4
   auxiliary-stream events graph-scoped lifetimes. This restores safe overlap
   for both MTP and DSpark and replaces global serialization workarounds.
9. #104 - native mixed
   NVFP4/MXFP4 plus NF3 MoE serving. Merge after the B12X NF3 geometry fix.
10. #109 - consolidated
    DSpark/DFlash correctness, optional FP8 draft head, and optional load-aware
    compact verification. Merge after #108; merging it after #97-#103 also
    gives the intended complete speculative-decoding stack.

#109 supersedes and closes #105, #106, and #107. Do not merge those three old
PRs separately. The previous mixed #88 integration branch also remains closed.

Changes already canonical

• vLLM #95, the NVFP4
  MLA outer-scale and FP8-RoPE integration, is already merged.
• TP6 virtual sharding for target and draft models is already present through
  d437a23cfd; closed #96 is superseded and must not be merged separately.
• B12X NF3 format support is already present in master through merged [codex] consolidate wiki navigation #31.

Integration verification completed

The open vLLM PRs above were merged locally in the stated dependency order.
The merge completed without conflicts.

• ruff check: passed for all 86 changed Python files.
• ruff format --check: passed for all 86 changed Python files.
• git diff --check: passed.
• Integrated v17-runtime test sweep: 366 passed, 714 skipped.
• One missing-attribute guard found by that sweep was fixed in #109 and its
  focused 15-test rerun passed.
• The PP2 scheduler test passed when two devices were exposed; its first run
  failed only because the test container exposed one GPU.
• The remaining test_scheduler_kv_connector_stats failure reproduces on
  unmodified current FF and is unrelated to this merge set.

Required post-merge validation

1. Create one fresh checkout of the merged FF head and one fresh checkout of
   the merged B12X master head.
2. Build the vLLM wheel and native extensions from those exact commits. Do not
   copy a different Python tree over an older wheel.
3. Record exact vLLM, B12X, FlashInfer, InstantTensor, PyTorch, CUDA, NCCL, and
   build-repository commits in the image labels and release documentation.
4. Confirm that generated version metadata reports the same vLLM commit whose
   Python sources are imported at runtime.
5. Validate GLM-5.2 TP8/DCP1 MTP-off and MTP:3, including acceptance and CC1
   decode. MTP:3 must retain the recovered fixed-width target-step rate.
6. Validate GLM-5.2 TP6 with MTP and DCP, A4/A16, online quantization, NVFP4 KV,
   and the NF3 hybrid checkpoint where applicable.
7. Validate DS4 TP2 and TP4 with MTP-off, MTP:2, and DSpark. Run long-context
   output checks in addition to throughput sweeps.
8. Validate A8 W4A8 capture with shared-expert auxiliary streams enabled; no
   stream-disabling workaround is acceptable.
9. Build and publish only after the canonical source image passes these checks.

Explicit exclusions

The following are not silently included in this merge set:

• upstream PR #47979 sequence-parallel/SM120 PCIe experiments;
• deployment launchers, Compose defaults, NUMA pinning, and memory envelopes;
• release-specific KV headroom or graph-size tuning;
• Docker filesystem overlays used only for A/B diagnosis.

If any excluded item is still required, it must receive its own PR in the
appropriate source or build repository before the release image is produced.

Completion criteria

• all listed canonical PRs are merged in dependency order;
• superseded PRs and release-only integration branches are clearly marked;
• a fresh source build contains no unpublished patch or copied source overlay;
• image provenance is reproducible from public commits and one build command;
• the post-merge correctness and performance matrix passes without hidden
  runtime switches.