upstream_watchlist.yaml

watch:
- upstream: vllm#39598
  status: closed          # network sweep 2026-07-05: PR CLOSED unmerged (was 'open'); the residual stays unmitigated upstream — keep the probe gate
  severity: P1-correctness
  local_patches:
  - P64
  action: a-b-test        # probe-on-our-stack (stream_toolcall_probe.py); enum-mapped
  since: 2026-07-02
  notes: |
    club-3090 #145 close-out (2026-06-13, read from issue comments): v0.22.0
    fixed tool_choice=auto + streaming natively, BUT a RESIDUAL remains on stock
    vLLM: tool_choice=REQUIRED + thinking + streaming DROPS 13/20 tool calls
    with MTP n=3 (clean 0/20 without MTP; parser-independent) - vllm#39598
    resurfacing un-mitigated. Genesis P64 was the mitigation and is version-
    capped off on >=dev491 pins (engine-native parser owns streaming). Our 35B
    runs MTP K=5 + qwen3_xml -> DIRECT exposure for tool agents that use
    tool_choice=required with streaming. Their 8-packs never caught it because
    packs run non-streaming; they built scripts/stream-toolcall-probe.py for it.
    PROBED 2026-07-02 on the live 35B (dev672, MTP K=5, qwen3_xml, 190W caps):
    scripts/stream_toolcall_probe.py --n 20 -> 0/20 dropped, 0 errors — the
    residual does NOT reproduce on our pin. Two turns showed a cosmetic
    content-channel XML fragment ALONGSIDE a correctly delivered tool_call
    delta (no loss; watch if agent frontends render stray XML). KEEP the probe
    as a pin-bump gate: the drop reproduced on club's stack/pins, and the
    fleet_boot_smoke tool-call check is auto/non-streaming so it cannot see it.
- upstream: vllm#40886
  status: open            # network sweep 2026-07-05 (audit 2026-07-04 #27): gh pr view says OPEN, NOT merged — the 'merged / absorbed-in-v0.24.0' claim was a club-3090 CHANGELOG assertion taken on faith (iron-rule-#11 class). Do NOT retire G4_04 until a verified merge (gh pr view) or a verified native equivalent in the target pin.
  severity: P2-cleanup
  local_patches:
  - G4_04
  action: watch           # was retire; retire is gated on VERIFIED merge (audit #27) — restore action=retire only with gh-pr-view proof
  since: 2026-07-02
  notes: |
    CORRECTION 2026-07-05: vllm#40886 is OPEN unmerged (gh pr view). The
    v0.24.0-absorption claim below is club-3090's, unverified — if their
    v0.24.0 ships the remap it came from another commit/PR; identify it
    before any G4_04 retire-audit.
    club-3090 shipped vllm-stable v0.24.0 "overlay-free" (their #533, 2026-06-30):
    upstream absorbed the Gemma-4 AWQ MoE keys remap (#40886) that G4_04 vendors.
    We are NOT on v0.24.0 yet (current pin dev672 = 0.23.1). ON A v0.24.0 BUMP,
    deep-diff G4_04 remap vs native #40886 (iron-rule #11): RETIRE if byte-equivalent,
    else KEEP the extra keys. Source: club-3090 CHANGELOG (v0.22.0 -> v0.24.0).

    ALSO ON THE v0.24.0 BUMP — further overlay-free retire-audit candidates club
    dropped (was a separate `v0.24.0-overlay-free` entry; folded here because none
    of these track a single mergeable PR — G4_25 is genesis_original/upstream_pr=None
    — so they cannot be schema-modeled as standalone PR entries). iron-rule #11: do
    NOT retire on a title match.
    (a) Gemma p-RoPE long-context: a v0.24.0 config fix may supersede G4_25 (Gemma 4
        dual-RoPE base-freq divergence guard) — but G4_25 is a genesis_original
        DIAGNOSTIC (warns on rope_theta == global_rope_theta), not a PR backport;
        keep unless the upstream config validation makes the warning unreachable.
    (b) truncate_prompt_tokens (#41800): native in v0.24.0; we vendor no equivalent.
    (c) marlin-pad native: audit our marlin-path patches (P37/P81 family).
    (d) club found MTP SLOWER on Gemma-4 MoE with a FIXED K -> they repointed their
        dual to MTP-off (#537). We do NOT ship a fixed K: SNDR_MTP_DYNAMIC_K_001
        (per-seq adaptive-K MTP proposer, port of vllm#26504 — rolling accept-rate
        window len=10 + hysteresis; K++ when acceptance high, K-- down to 1 when low)
        + PN40-C (adaptive K) + per-task K presets (gemma4-*-mtp-chat-k3 /
        -structured-k4). Our adaptive-K lowers K exactly on the low-acceptance tasks
        where club's fixed-K lost -> their finding VALIDATES our design. Action:
        confirm adaptive-K is enabled/tuned on the gemma-4 MTP presets (per-task,
        acceptance-driven) — NOT disable MTP wholesale.
    (e) W8A8 int8 (weights+activations) ~zero quality cost, +17-51% prefill,
        TTFT 122 vs 158 ms (their experimental v0.24.0 row) -> a prefill-optimised
        variant worth an A/B if we add one.
    Source: club-3090 CHANGELOG + BENCHMARKS (2026-06-27..07-01).
- upstream: vllm#43409
  status: merged
  fix_upstream: vllm#45656
  fix_status: merged-in-pin  # network sweep 2026-07-05: #45656 (merged 06-18) is an ancestor of BOTH live pin bases (dev714/dev748); pristine dev748 auto_gptq.py:761-769 carries the is_sym use_zp guard natively
  severity: P0-correctness
  local_patches:
  - PN400
  action: drift-check  # reconciled 2026-06-23: PN400 IS the backport of fix_upstream vllm#45656 (port complete); drift-check until #45656 lands in our pin, then retire PN400. DONE: PN400 retired 2026-06-24 (dev301 bump); row kept for provenance.
  since: 2026-06-20
  notes: |
    REGRESSION IN dev148 (found by the /loop upstream sweep 2026-06-20).
    #43409 (merged 06-12, IN dev148) removed the `if not is_sym` qzeros
    guard in AutoGPTQMoEMethod.get_fused_moe_quant_config AND
    CompressedTensorsWNA16MarlinMoEMethod.process_weights_after_loading.
    Symmetric (is_sym=True) AutoRound/GPTQ Marlin MoE on NVIDIA then passes
    meaningless qzeros to the Marlin kernel -> INCORRECT MoE expert outputs.
    Our 27B = Qwen3.6-27B-int4-AutoRound is CONFIRMED on the broken path:
    checkpoint config.json quantization_config = {quant_method: auto-round,
    sym: True, bits: 4, group_size: 128} on NVIDIA A5000 (verified live
    2026-06-20). Fix #45656 ("Restore is_sym guard for zp in GPTQ/CT MoE",
    merged 06-18 16:20Z) landed ~12h AFTER the dev148 base commit
    (b4c80ec0f @ 06-18 04:18Z) -> NOT in pin.
    ACTION: backport #45656 as a Genesis TextPatch (2-file additive: restore
    `use_zp = not is_sym` gating of w1_zp/w2_zp in auto_gptq.py + the
    `if w13/w2_qzeros is not None` guards in
    compressed_tensors_moe_wna16_marlin.py), version-capped to pins lacking
    #45656; OR wait for the next pin bump including #45656 and treat dev148
    as not-27B-correctness-safe until then. VALIDATE via 27B greedy-output
    A/B (dev148 vs dev148+patch) on a deterministic prompt. The 27B is not
    currently running (35B PROD is up) so there is no live mis-serving now.
- upstream: vllm#43559
  status: merged
  fix_upstream: vllm#43650
  fix_status: open-NOT-in-pin
  severity: P0-correctness
  local_patches: []
  affects_models:
  - Qwen3.6-27B-int4-AutoRound
  - Qwen3.6-35B-A3B-FP8
  action: watch
  since: 2026-06-21
  notes: |
    REGRESSION in dev148 (found by the /loop upstream sweep 2026-06-21).
    MTP + --enable-prefix-caching on hybrid GDN silently drops accuracy
    ~20% (tool-eval 90%->50%), confirmed upstream on BOTH Qwen3.6-27B and
    35B-A3B. Fix #43650 is OPEN, NOT in dev148. RIG STATUS: 35B PROD is
    SAFE — start_qwen3.6-35b-balanced.sh runs MTP K=5 but has NO
    --enable-prefix-caching (verified live 2026-06-21). 27B INT4 is exposed
    only if launched with MTP + prefix caching. ACTION: never combine MTP +
    --enable-prefix-caching on hybrid GDN models; if prefix caching is
    required, force --kv-cache-dtype fp8 (issue-confirmed workaround). When
    #43650 merges, take the final-mamba-block-drop fix. Gemma-4-26B-A4B is
    NOT hybrid -> unaffected.
- upstream: vllm#46384
  status: open
  severity: P0-correctness
  local_patches:
  - PN346
  - PN346B
  affects_models:
  - Qwen3.6-35B-A3B-FP8
  - Qwen3.6-27B-int4-AutoRound
  action: drift-check
  since: 2026-06-26
  notes: |
    INCOMING ANCHOR-DRIFT (deep re-study 2026-06-26, pin dev424
    0.23.1rc1.dev424+g3f5a1e173). #46384 ("[2/N] partial prefix-cache hits in
    hybrid coordinator") is an UNCONDITIONAL structural rewrite of
    find_longest_cache_hit in BOTH v1/core/kv_cache_coordinator.py AND
    v1/core/single_type_kv_cache_manager.py — the exact files+regions PN346
    (manager half, vendor #43650) and PN346B (coordinator curr_hit_length
    clamp #45614 + Part-B FA/Mamba post-trim belt #46281) anchor into with
    required=True sub-patches. NOT a runtime bug for us today: the partial-hash
    DATA path is gated on enable_partial_hash_hits (needs
    mamba_cache_mode="align"; PROD runs "none" -> inert). BUT the rewrite is
    mode-independent: it deletes the _get_block_hashes closure, reshapes the
    clamp region (_new_hit_length -> get_cache_hit_length, cdiv truncation,
    KVCacheBlockListWithHitLength) and edits the MambaManager loop region. On
    the pin that carries #46384 the PN346/PN346B required anchors VANISH -> the
    transactional patcher SKIPs -> SILENT LOSS of the #43559 Mamba+EAGLE/MTP+APC
    poison guard (the exact 35B-A3B MTP+APC path #43559 targets). The anchor-SOT
    bump_preflight catches this mechanically (PN346/PN346B land in
    genuine_anchor_drift on the carrying pin); this entry records the PLAN.
    ACTION on merge: re-anchor PN346 onto the new MambaManager partial branch;
    for PN346B verify whether the merged clamp (get_cache_hit_length + cdiv)
    already encodes the monotonic-min — if so RETIRE PN346B Part-A, else
    RE-ANCHOR onto the new clamp shape; re-derive the Part-B belt against the
    cdiv truncation. Do NOT vendor #46384 itself (half-merged [2/N] partial-hash
    machine, inert on mamba_cache_mode="none"). Cross-ref sweep: pr 46384.
- upstream: vllm#46446
  status: open
  severity: P3-perf
  local_patches:
  - PN129
  affects_models:
  - Qwen3.6-35B-A3B-FP8
  - Qwen3.6-27B-int4-AutoRound
  action: retire
  since: 2026-06-26
  notes: |
    RETIRE-ON-MERGE (deep re-study 2026-06-26). #46446 ("Warm up slot-mapping
    kernel through BlockTable") adds BlockTable.warmup_compute_slot_mapping()
    (+ MultiGroupBlockTable fan-out), called from
    gpu_worker.compile_or_warm_up_model before cudagraph capture — the CLEAN
    upstream of PN129 (both vendor #42165). PN129 monkeypatches
    Worker.compile_or_warm_up_model and pokes the private Triton
    do_not_specialize attribute (self-documented hacky); #46446 does it the
    supported way (real compute_slot_mapping over multiple shapes via
    BlockTable, no private-API poke) and runs for MultiGroupBlockTable which our
    hybrid (FA+GDN+Mamba) uses. PN129 is default-OFF and NOT enabled on prod
    YAMLs -> NO correctness exposure (boot-JIT latency only). ACTION on merge:
    retire PN129 entirely. OPTIONAL before merge: re-base PN129 onto #46446's
    warmup_compute_slot_mapping (drop the private-Triton hack). Not urgent.
    Cross-ref sweep: pr 46446.
- upstream: vllm#38999
  status: open
  severity: P1-crash
  local_patches: []
  affects_models:
  - Gemma-4-26B-A4B
  action: watch
  since: 2026-06-21
  notes: |
    Gemma-4-26B-A4B CRASHES with --data-parallel-size>1 (AssertionError in
    cuda_communicator all_gather on the FIRST request — loads fine, graphs
    build, /health returns 200, then dies in serving). Found by the /loop
    sweep 2026-06-21. TRAP: setting DP=2 for "parallelism" = a silent
    in-flight incident after a healthy boot. ACTION: never set
    --data-parallel-size>1 or --enable-eplb on Gemma4. The validated speedup
    path is --enable-expert-parallel (EP-over-TP) — rig-validated -20% TPOT
    (journal §78), now canonically emitted by compose() for gemma4_moe
    multi-GPU. Fresh EP fixes #43219/#43110/#45002 do not regress that path.
- upstream: vllm#44740
  status: open
  fix_upstream: vllm#44209
  fix_status: open
  severity: P1-crash
  local_patches: []
  affects_models:
  - Qwen3.6-27B-int4-AutoRound
  action: watch
  since: 2026-06-21
  notes: |
    MTP cudagraph negative memory estimate (-35 GiB) -> KV reallocation ->
    OOM. On hybrid GDN the KV pool size is non-deterministic -> a silent
    crash-loop under restart:unless-stopped AFTER /health already returned
    200. Found by the /loop sweep 2026-06-21. Affects our 27B INT4 (MTP +
    hybrid GDN + cudagraph). WORKAROUND: pass an explicit
    --kv-cache-memory-bytes instead of pure --gpu-memory-utilization; do NOT
    treat the healthcheck as a success indicator when MTP is on.
- upstream: vllm#38479
  status: merged
  local_patches:
  - P3
  - P4
  - P5
  - P6
  - P22
  - P26
  - P28
  - P36
  - P38
  - P44
  - P46
  action: a-b-test
  since: 2026-04-15
  notes: 'Upstream merged TurboQuant KV cache compression. Genesis stack

    (P3..P46) overlaps significantly. Need A/B vs upstream''s

    `--kv-cache-dtype turboquant_k8v4` to confirm Genesis advantage

    on hybrid (Mamba/DeltaNet) models that upstream excludes per

    §10.10 audit note.

    '
- upstream: vllm#40914
  status: open
  local_patches:
  - P67
  - P67b
  action: a-b-test
  since: 2026-05-08
  notes: 'Routes uniform K+1 spec-verify batches through the multi-query

    kernel. P67/P67b are Genesis''s own implementation. After upstream

    merge: A/B comparison required — keep Genesis path if it

    outperforms (likely so on Ampere; uncertain on Hopper+).

    '
- upstream: vllm#40069
  status: open
  local_patches:
  - P4
  - P5
  - P67
  - P67b
  - P78
  - P98
  - P99
  - P101
  - PN57
  action: drift-check
  since: 2026-05-08
  notes: 'TurboQuant/HIGGS attention follow-up tracking issue. Genesis stack

    should auto-retire patches as their corresponding bullet points

    land in upstream PRs.

    '
- upstream: vllm#43432
  status: closed          # network sweep 2026-07-05: PR CLOSED unmerged (was 'open'); the value-MSE quality idea remains un-landed upstream — do-not-auto-adopt guidance below stands
  local_patches:
  - P44
  - P67
  - PN119
  - P22
  - P26
  - P32
  - PN261
  action: a-b-test
  since: 2026-06-19
  notes: '[TurboQuant] Use MSE (Lloyd-Max) quantization for VALUES — a QUALITY

    candidate; the only non-applicable-condemned delta of the delta-scan.

    NOT condemned by the FIX2 dead-end: FIX2 was a decode-THROUGHPUT

    MSE-KEY tensor-core experiment that died at Gemma GQA group=2 (needs

    M>=16); THIS is scalar-Triton VALUE-quant accuracy, independent of the

    GQA ratio. PR benches our exact 27B-int4-AutoRound 36/36 NIAH; Gemma

    V-MSE mean NLL 0.297 vs 0.607 uniform + 2B/value-vector smaller slot.

    DO NOT auto-adopt: OPEN/bot-review-only, 1124-line rewrite of value

    store/decode/dequant on the live k8v4 hot path BOTH PROD models run,

    changes value_packed_size 68->66 / slot_size 196->194 (must co-version

    PN261/PN119/P22/P26/P32), extra per-decode inverse-Hadamard@Pi GEMM

    unquantified for k8v4 TPOT. If pursued: backport ONLY the value-MSE

    half (drop bundled mm_prefix_range — already vendored in

    overlays/pr42637), default-OFF PNxxx extending overlays/pr42637

    store+decode, with a mandatory dev148 rig A/B (TPOT + tool-call

    clean-rate + NIAH on 27B & 35B) BEFORE enabling.

    '
- upstream: vllm#45748
  status: open
  local_patches: []
  action: watch
  since: 2026-06-19
  notes: 'Turboquant native fp8 v4 (CUDA) store. NOT APPLICABLE — hardware-gated

    SM>=8.9 (Hopper/Ada E4M3). Our 2x RTX A5000 = SM (8,6) deterministically

    take the fp8e4b15 branch (live store, Genesis P3 workaround), which is

    byte-incompatible with the native kernel''s E4M3; the native op is not

    even compiled into dev148 (torch.ops._C.turboquant_store_fp8_v4 = False).

    Zero perf/correctness delta. Revisit ONLY on a Hopper/L40 migration; if

    ever ported, preserve the fp8_e4b15==0 / device_capability>=89 guards

    VERBATIM or it silently corrupts the A5000 KV cache. Do NOT re-flag as

    a missed backport.

    '
- upstream: vllm#41803
  status: open
  local_patches: []
  action: watch
  since: 2026-06-19
  notes: '[Attention][MLA] Triton-fused TurboQuant decode backend. NOT APPLICABLE

    — MLA-only. Our models are GQA/hybrid, not DeepSeek-MLA: 35B =

    qwen3_5_moe (n_kv_heads=2, head_dim=256, no kv_lora_rank), 27B = DFlash

    GDN+Mamba hybrid. The backend enters only via the MLA priority list +

    the kv_lora_rank-gated spec hook, never touches our GQA decode kernel

    (P67/P40/PN119). Do NOT re-flag as a missed backport despite the

    "TurboQuant decode" title match. Revisit only if a DeepSeek-MLA model

    joins the fleet (then with vllm#38479).

    '
- upstream: vllm#43887
  status: closed          # network sweep 2026-07-05: PR CLOSED unmerged (was 'open'); the consciously-declined assessment below stands, drift risk gone with the closure
  local_patches:
  - P67b
  - G4_67
  - G4_81
  action: drift-check
  since: 2026-06-19
  notes: '[TurboQuant] Add MTP spec-decode routing. ALREADY COVERED live by P67b

    (+32% TPS) with G4_67 (our #40914 backport) and G4_81 as backups. The

    PR flips supports_spec_as_decode=True — the OPPOSITE of our deliberate

    design (live=False); G4_81 checklist item 5 documents that flip as a

    silent-corruption + PN243-empty-max hazard. Its only NEW half

    (value-MSE) comes from base #43432, tracked there. Assessed and

    consciously declined. IF it merges: verify G4_81 wrap-outer

    interception still matches the new upstream forward() drift markers

    (no double-route).

    '
- upstream: vllm#43878
  status: closed          # network sweep 2026-07-05: PR CLOSED unmerged (was 'open'); P101 mitigation remains our path
  local_patches:
  - P101
  - PN116
  action: drift-check
  since: 2026-06-19
  notes: '[TurboQuant] streaming fallback for long prefill. NOT a miss — the

    residual 0.5GB FP16 continuation-alloc spike that could affect us is

    already mitigated differently by P101 (cached_len>=32768 -> sliced

    decode-replay). The PR streaming core is gated behind

    `not _can_use_flash_prefill`, unreachable on our flash-enabled

    head_dim=256 / sliding_window=None / text-only PROD path -> would be

    600+ dormant LOC. PR self-flags "could not validate k8v4 serving" =

    our exact cache layout.

    '
- upstream: vllm#40269
  status: merged          # network sweep 2026-07-05 (audit 2026-07-04 #28): MERGED 2026-05-14 — sat hidden as 'open' for 7 weeks because the watchlist audit is offline-only
  local_patches:
  - P77
  - PN40
  action: a-b-test        # was port; nothing left to port — the merge is an ancestor of the live pin bases, so the code is IN dev714/dev748. Remaining task is the enable/verify probe described in the notes
  since: 2026-05-08
  notes: 'Probabilistic draft rejection — passes draft_probs from drafter

    to verifier. Genesis P77 (adaptive ngram K) is orthogonal but

    complementary. Port as PN90 when upstream merges.

    UPDATE 2026-07-05: MERGED 2026-05-14 and an ancestor of the live pin

    bases — the draft_probs wiring is IN dev714/dev748 natively. The

    "port" here is now an ENABLE/VERIFY task, not a backport: check

    whether our MTP drafters populate draft_probs on the live pin and

    whether the probabilistic rejection path engages (or needs a config

    knob) before writing any patch. Do NOT vendor a copy of merged code.

    '
- upstream: vllm#26504
  status: closed          # network sweep 2026-07-05: PR CLOSED unmerged (was 'open'); the MTP-adapt design sketch below remains the reference if we ever revive it (SNDR_MTP_DYNAMIC_K_001, our port, is itself retired)
  local_patches: []
  action: port
  since: 2026-05-31
  notes: "DynamicProposer — per-sequence adaptive K based on rolling\nacceptance-rate window. Claims +5-12%\
    \ TPS on mixed workload\nvs static K. **Limitation for Genesis**: PR extends\n`EagleProposer` directly,\
    \ NOT MTP-compatible as-written. Our\n4 PROD models use MTP K=3/K=4 (not Eagle). Backport requires\n\
    MTP adaptation: either inherit from MTP-proposer instead of\nEagle, or extract a mixin/composition\
    \ pattern. STUDY done\n2026-05-31 (see commit a074e900 follow-up notes). PR has\n`needs-rebase` label\
    \ — wait for author to rebase before\nattempting the MTP-adapt backport. Acceptance threshold +\n\
    hysteresis algorithm is generic enough to lift.\n\nMTP-adapt design sketch (2026-05-31 study, ready\
    \ for vendor\nsession when upstream rebases):\n\nArchitecture target — extend DraftModelProposer instead\
    \ of\nEagleProposer. The DraftModelProposer (at\nvllm/v1/spec_decode/draft_model.py in pin 0.21.1rc1.dev354+g626fa9bba)\n\
    is the assistant-model-based MTP proposer all 4 of our PROD\ngemma4/qwen3.6 PROD models go through.\
    \ It inherits from\nSpecDecodeBaseProposer which already owns\n`self.num_speculative_tokens` — the\
    \ same attribute the\nDynamicProposer mutates per-batch.\n\nAlgorithm port (1:1 with #26504 file):\n\
    \  1. Per-seq SequenceState (acceptance_rate_history deque\n     of len 10, num_spec_tokens int).\n\
    \  2. ACCEPTANCE_HISTORY_LEN=10, ACCEPTANCE_RATE_HYSTERESIS=0.05,\n     MIN_HISTORY_FOR_ADJUSTMENT=3,\
    \ MIN_SPEC_TOKENS=1 — copy\n     the constants.\n  3. In propose() override:\n     a. Update sequence\
    \ states from prev step's\n        `self.runner.input_batch.num_accepted_tokens_cpu`.\n     b. Cleanup\
    \ states for finished req_ids.\n     c. Compute per_sequence_k via\n        _adjust_and_get_spec_tokens_for_batch().\n\
    \     d. Save self.num_speculative_tokens, set to\n        max(per_seq_k), call super().propose(),\
    \ restore.\n     e. Pad the returned tensor to self.max_spec_tokens\n        columns so the runner\
    \ expects a fixed shape.\n\nConfig plumbing:\n  - Add SpecConfig field `acceptance_rate_threshold:\
    \ float =\n    0.7` (PR's default).\n  - Add method-name route: `--speculative-config\n    '{\"method\"\
    : \"mtp_dynamic\", \"num_speculative_tokens\": 4,\n      \"acceptance_rate_threshold\": 0.7, ...}'`\n\
    \  - SpecConfig.proposer_class resolution must dispatch\n    mtp_dynamic → Genesis MTPDynamicProposer.\n\
    \nImplementation surface (estimated 1-2 days):\n  - vllm/sndr_core/integrations/spec_decode/mtp_dynamic_proposer.py\n\
    \    (NEW, ~250 LOC mirroring PR #26504's structure with\n    DraftModelProposer base).\n  - vllm/sndr_core/integrations/spec_decode/spec_decode_config_routing.py\n\
    \    patch — add mtp_dynamic to method enum + dispatch.\n  - registry.py entry — PN_NEW with method=mtp_dynamic\
    \ gate.\n  - Test on gemma4-31B (4-class workload) + qwen3.6-27B (chat\n    + structured). Expected\
    \ +5-12% per the PR's claim; the\n    actual gain on our hybrid GDN+Mamba (27B) and MoE A3B\n    (35B)\
    \ stacks would be the first empirical check.\n\nOperator value: 31B chat-K=3 + structured-K=4 split\
    \ currently\nrequires the spec_decode gateway to route per request. A\nmtp_dynamic proposer would\
    \ converge a SINGLE launcher to the\nright K per-sequence automatically — no gateway needed for\n\
    the K-axis. The chat/structured workload-class semantic split\nstays valid (different compression_plan\
    \ / drafter behavior),\nbut the K choice within each profile becomes self-adapting.\n"
- upstream: vllm#42006
  status: closed          # network sweep 2026-07-05: PR CLOSED unmerged (was 'open'); the retire-on-merge trigger can never fire from THIS PR — keep the overlays, superseded locally by the #42237 v2 vendor anyway
  local_patches:
  - G4_T1
  action: retire
  since: 2026-05-30
  notes: 'Gemma4 MTP streaming multi-tool-calls fix — segment-replay

    refactor of `_extract_streaming`. Vendored as G4_T1 v1 overlay

    `g4_t1_gemma4_tool_parser_pr42006_overlay.py` (matches

    noonghunna/club-3090 stack 2026-05-08). On merge: delete both

    v1 and v2 overlay files + remove `-v` mount in launcher

    `start_gemma4-31b-tq-mtp-structured-k4.sh`. SUPERSEDED locally by

    the v2 vendor of PR #42237 (see next entry) but kept here for

    retire-on-merge symmetry.

    '
- upstream: vllm#42237
  status: closed          # network sweep 2026-07-05: PR CLOSED unmerged (was 'open'); retire-on-merge cannot fire — the G4_T1 v2 overlay stays REQUIRED for gemma4 streaming tool calls until an equivalent lands upstream
  local_patches:
  - G4_T1
  action: retire
  since: 2026-05-31
  notes: 'Gemma4 streaming tool parser REWRITE in Hermes/Kimi style

    (accumulated text + rescan + diff). Net -103 LOC vs PR #42006

    and drops the third-party `regex` import. Currently vendored

    as G4_T1 v2 overlay

    `g4_t1_v2_gemma4_tool_parser_pr42237_overlay.py` — this is

    the file the launcher bind-mounts as of 2026-05-31. Empirical

    result on gemma4-31B AWQ-4bit + TQ k8v4 + MTP K=4: 35/35 tool

    calls (7 cases × 5 runs, both streaming and non-streaming,

    with `Connection: close` header — see Bug Class 13 notes in

    docs/TROUBLESHOOTING.md). On merge: delete both overlay files

    + remove launcher `-v` mount. The HTTP keep-alive parser

    state-leak (case 5 nested-object missing closing `"}` two

    chars on runs 2..N when client uses keep-alive) is a residual

    vLLM-side bug worth flagging to the PR author when the merge

    is being prepared.

    '
- upstream: vllm#43074
  status: closed          # network sweep 2026-07-05: PR CLOSED unmerged (was 'open'); retire-on-merge cannot fire — Q3_T1 v1 stays live; watch the related PR list below for a successor fix
  local_patches:
  - Q3_T1
  action: retire
  since: 2026-05-31
  notes: "[Bugfix] Qwen3Coder streaming: emit args when whole tool body\nlands in one delta. PR #43074\
    \ fixes the EXACT bug class our\nQ3_T1 v1 overlay addresses — the upstream state machine drops\ntool_calls\
    \ when MTP K=3 packs the whole `<function=NAME>...\n</function></tool_call>` body into a single delta\
    \ after\nis_tool_call_started flipped on the prior delta. Reproduces\nwith `--tool-call-parser qwen3_coder\
    \ --stream-interval 20\n--speculative-config '{\"method\":\"mtp\",\"num_speculative_tokens\":3}'`\n\
    on Qwen3.6-27B — exactly our 27B PROD launcher config.\n\nAuthor's fix is a surgical +81/-41 LOC patch\
    \ to the state\nmachine (\"collapsed body\" detection + emit args inline).\nOur Q3_T1 v1 takes a different\
    \ approach — full Hermes/Kimi\nrescan + diff rewrite (~250 LOC of new method) — which is\nmore general\
    \ (covers ALL packing patterns, not just \"whole\nbody in one delta\") but heavier. After upstream\
    \ merges,\nverify the fix covers our 7-case bench at 100% on both 27B\nand 35B, then retire Q3_T1\
    \ v1.\n\nRelated (track but not primary retire triggers):\n  - vllm#40861 (Qwen3 XML and Coder streaming\
    \ regression\n    fix, needs-rebase)\n  - vllm#39598 (streaming tool call empty fields with MTP\n\
    \    + qwen3coder early return, needs-rebase)\n  - vllm#41970 (Enable tool parser tests for Qwen3\
    \ Coder)\n  - vllm#37829 (stop string streaming at next <parameter=>\n    boundary, needs-rebase)\n\
    \  - vllm#43363 (refactor Use shared type coercion utilities\n    in Qwen3 XML — touches type-coercion\
    \ path our Q3_T1 v1\n    also depends on via _convert_param_value)\n\nEmpirical state of Q3_T1 v1\
    \ on pin 0.21.1rc1.dev354+g626fa9bba:\n  qwen3.6-27b stream: 35/35 ✓ (was 1/7 = 14%)\n  qwen3.6-35b\
    \ stream: 35/35 ✓ (was 2/14 = 14%)\n  non-streaming on both: unchanged 7/7\nSee Bug Class 15 in docs/TROUBLESHOOTING.md\
    \ for full bench.\n"
- upstream: vllm#43909
  status: merged
  local_patches: []
  action: a-b-test
  since: 2026-05-31
  notes: "[Bug] Fix gemma4 MTP IMA issue when TP>1 (CUDA error: illegal\nmemory access). Merged 2026-05-30\
    \ (after our pin 626fa9bb of\n2026-05-28). Minimal +9/-1 LOC fix in\n`vllm/model_executor/models/gemma4_mtp.py`:\
    \ caches\n`_stable_full_lm_head_weight` to prevent CUDA graph from\ncapturing a dangling pointer to\
    \ a temp tensor (post all_gather\n+ slice). Author bench on Gemma-4-E2B-it shows 0.64s avg\nlatency\
    \ for batch=16 input-len=512 output-len=128.\n\nVendored 2026-05-31 as Genesis G4_77 overlay (4-file\n\
    bind-mount to `/usr/local/lib/python3.12/dist-packages/\nvllm/model_executor/models/gemma4_mtp.py`).\
    \ A/B bench on\ngemma4-31b AWQ + TQ k8v4 + MTP K=4 + TP=2 + 2× A5000 +\ngpu-mem-util=0.80 (Bug Class\
    \ 14 fix), 3 samples × 4\nworkloads, greedy decode, `Connection: close`:\n\n  | Workload | Baseline\
    \ | With G4_77 | Δ |\n  |---|---:|---:|---:|\n  | chat | 23.54 | 20.27 | -13.9% |\n  | code | 41.73\
    \ | 42.95 | +2.9% |\n  | count | 52.37 | 50.05 | -4.4% |\n  | json | 53.47 | 47.63 | -10.9% |\n\n\
    NET NEGATIVE. Reverted. Likely cause: the cached tensor\n(post all_gather across TP=2) consumes additional\
    \ VRAM at\na config that is already running gpu-mem-util=0.80 to give\nGenesis P38 TQ workspace its\
    \ 1 GiB alloc. Reduced KV cache\n→ more KV churn on longer prompts → measurable regression\non chat\
    \ (longest output) and JSON (structured but long).\nThe PR's author bench was on Gemma-4-E2B-it (much\
    \ smaller\nmodel, presumably more memory headroom). The IMA workaround\nGenesis already ships (G4_71..G4_76\
    \ stack in launcher) may\nalso be incompatible with the upstream cache in subtle ways.\nRe-evaluate\
    \ after next pin bump if the upstream lands in\nproduction builds.\n"
- upstream: vllm#43803
  status: merged
  local_patches: []
  action: a-b-test
  since: 2026-05-31
  notes: "[Perf] remove seqlen from Mamba SSD chunk kernels. Merged\n2026-05-28. +0/-10 LOC — removes\
    \ a dead `seqlen` parameter\nfrom 5 Mamba prefill kernels (`_chunk_cumsum_fwd_kernel`,\n`_chunk_state_fwd_kernel`,\
    \ `_chunk_scan_fwd_kernel`,\n`_bmm_chunk_fwd_kernel`, `_causal_conv1d_fwd_kernel`).\nTriton specializes\
    \ plain `int` args by default, so every\nprefill with a new total-token-count was triggering a\nfresh\
    \ JIT compile. Author claim on Nemotron-H-8B / A10G /\nShareGPT: Mean TTFT -17%, P99 TTFT -46%. Throughput\n\
    unchanged.\n\nVendored 2026-05-31 to qwen3.6-27b-int4-autoround\nlauncher as a 4-file bind-mount (`/tmp/mamba_fixed/\n\
    {causal_conv1d,ssd_bmm,ssd_chunk_scan,ssd_chunk_state}.py`).\nCleared ~/.triton/cache before measurement\
    \ to force fresh\nJIT path. A/B vs baseline 27B, 8-sample TTFT × 3 prompt\nlengths, 3-sample TPS ×\
    \ 4 workloads, greedy, `Connection:\nclose`:\n\n  | Metric | Baseline | With #43803 | Δ |\n  |---|---:|---:|---:|\n\
    \  | TTFT short (5 tok prompt) | 151.3ms | 139.4ms | -7.9% |\n  | TTFT medium (80 tok prompt) | 169.7ms\
    \ | 176.0ms | +3.7% (noise) |\n  | TTFT long (400 tok prompt) | 233.2ms | 232.5ms | -0.3% (noise)\
    \ |\n  | TPS chat | 91.68 | 86.95 | -5.2% |\n  | TPS code | 113.53 | 113.05 | -0.4% (noise) |\n  |\
    \ TPS count | 117.79 | 116.43 | -1.2% (noise) |\n  | TPS json | 105.61 | 107.77 | +2.0% |\n\nNET marginal-to-negative.\
    \ Reverted. Does NOT reproduce the\nPR's claimed -17%/-46% on our config. Likely cause: our\nqwen3.6-27B\
    \ is a hybrid GDN + Mamba architecture, NOT pure\nMamba like Nemotron-H. The GDN layers dominate the\
    \ prefill\nhot path and the Mamba SSD kernels see fewer invocations\nwith varying seqlen — fewer JIT\
    \ recompiles to save. The\nTPS chat regression (-5.2%) is just outside the run-to-run\nnoise band\
    \ (~3%) and may reflect a different code path\ntaken without the dead `seqlen` arg in our specific\n\
    configuration. Re-evaluate if Genesis later vendors a pure\nMamba model (e.g. nemotron-h-* family)\
    \ where this PR's\narchitecture would actually apply.\n"
- upstream: vllm#36138
  status: open            # network sweep 2026-07-05: gh pr view says OPEN, NOT merged (was 'merged' since 2026-04-25 — same claimed-merged-on-faith class as #40886, audit #27/#28)
  local_patches:
  - P59
  - P62
  - P64
  action: drift-check
  since: 2026-04-25
  notes: 'Grammar ignored when reasoning ended in speculated tokens.

    Already integrated as P62 + structured-output fixes. Verify

    anchor md5 not drifted after upstream pin bump.

    CORRECTION 2026-07-05: the PR is OPEN unmerged (gh pr view) — the

    original "merged" status was never verified. P62 remains OUR fix,

    not a backport-of-merged; do not treat it as a retire candidate.

    '
- upstream: vllm#40270
  status: open
  local_patches: []
  action: port
  since: 2026-05-08
  notes: 'Pluggable GPU KV cache eviction (LRU/2Q/ARC). Critical for

    agent/RAG workloads. Port as PN91 when merged.

    '
- upstream: vllm#37160
  status: merged          # network sweep 2026-07-05 (audit 2026-07-04 #28): MERGED 2026-04-01 — sat hidden as 'open' for 3 months; the second merged action=port row the offline-only audit missed
  local_patches: []
  action: watch           # was port; nothing left to port — the live pins ship the native kv-offload stack. Remaining task is an evaluation on non-hybrid lanes (see notes)
  since: 2026-05-08
  notes: 'CPU KV offload connector. Compatibility: NOT compatible with

    hybrid Mamba (Qwen3.5/3.6 GDN). Port as PN92 with `applies_to.is_hybrid: [False]`

    hard guard.

    UPDATE 2026-07-05: MERGED 2026-04-01; the live pins ship a native

    kv-offload stack (vllm/v1/kv_offload/ package + the offloading

    connector PN383 was written against). Nothing to port — the task is

    now an EVALUATION of native CPU KV offload on the non-hybrid lanes

    (Gemma-4), still with the is_hybrid guard rationale. Note the PN92

    name mentioned below was later used by an unrelated patch.

    '
- upstream: vllm#37190
  status: open
  local_patches: []
  action: watch
  since: 2026-05-08
  notes: 'MoE expert CPU offloading. Allow fitting 35B-A3B (256 experts)

    on single 24GB. High throughput cost (30-60% TPS hit). Port as

    experimental-only PN93.

    '
- upstream: vllm#38330
  status: open
  local_patches: []
  action: port
  since: 2026-05-08
  notes: 'Multimodal encoder cache. Cache vision embeddings across

    requests. Port as PN94 with preprocessor-version key.

    '
- upstream: vllm#37951
  status: open
  local_patches: []
  action: watch
  since: 2026-05-08
  notes: 'KV cache memory error at 262K context. Padding waste in hybrid

    layers. Genesis memory explain (DA-018) should surface this.

    '
- upstream: vllm#39998
  status: open
  local_patches:
  - P14
  action: watch
  since: 2026-05-08
  notes: 'Qwen3 + TurboQuant 4-bit KV + high concurrency + 32K context

    index out of bounds. P14 (block_table tail zero-fill) related;

    verify it covers the index path upstream is fixing.

    '
- upstream: vllm#40281
  status: closed
  local_patches: []
  action: watch
  since: 2026-05-08
  notes: 'Gemma4 support — closed unmerged. Wait for replacement PR.

    When Gemma4 lands upstream, create patch family + preset.

    '
- upstream: vllm#39133
  status: open
  local_patches: []
  action: watch
  since: 2026-05-08
  notes: 'Gemma 4 31B INT4 KV/cache limits on 2x24GB. SWA underutilized.

    Will need Gemma-specific config profile (memory estimator).

    '
- upstream: vllm#40265
  status: closed
  local_patches:
  - P7
  - P7b
  action: drift-check
  since: 2026-05-08
  notes: 'FLA TP device index overflow — closed unmerged. Pattern matters.

    Add regression guard in P7/P7b ensuring int64 device-index dtype.

    '
- upstream: Sandermage/sndr_core_engine#1
  status: open
  local_patches: []
  action: cookbook
  since: 2026-05-08
  notes: 'OOM with long context. See docs/COOKBOOK.md §1. Memory explain

    (DA-018) should preflight detect and warn.

    '
- upstream: vllm#40124
  status: open
  local_patches:
  - P3
  - P4
  - P5
  - P6
  - P14
  - P15
  - P22
  - P26
  - P28
  - P31
  - P36
  - P38
  - P44
  - P46
  - P67
  - P67b
  - P95
  action: drift-check
  since: 2026-05-08
  notes: 'Sander''s own master tracking issue: Ampere FP8 + TurboQuant +

    Qwen3.6-35B-A3B + hybrid MoE. Every Genesis patch in this list

    should map to a sub-bullet of the issue.

    '
- upstream: vllm#42102
  status: closed          # network sweep 2026-07-05: PR CLOSED unmerged (was 'open'); the independent-KV-groups backport plan below is not obsoleted by any merge — revisit only if DFlash lanes leave the archive
  local_patches:
  - PN21
  - PN23
  - PN24
  - PN38
  - PN40
  action: port
  since: 2026-05-12
  notes: "DFlash drafter coexistence with quantized target KV. Upstream\nproposes three mechanisms:\n\
    \  1. partition DFlash draft KV specs before page-size unify\n     (draft and target live in separate\
    \ KV groups).\n  2. drafter cache_dtype=\"auto\" when target global KV dtype is\n     quantized (drafter\
    \ inherits fp16/bf16 instead).\n  3. per-spec dtype in FlashAttention metadata scheduler.\n\nGenesis\
    \ already has PN21..PN40 for DFlash on Ampere. The relevant\nbackport target is the hybrid path where\
    \ target KV =\nturboquant_k8v4 while the drafter keeps fp16 (head_size=256\nconstraint).\n\nBackport\
    \ plan: see docs/_internal/research/upstream_42102_dflash_\nindependent_kv_groups_plan_2026-05-12.md\
    \ (S4.1 closure).\n\nCritical compatibility check: refined COMPAT-001 (qwen-next only)\nfrees Qwen3.6-27B\
    \ Lorbus + DFlash for further development.\n"
- upstream: vllm#38864
  status: open
  local_patches:
  - PN66
  - P12
  - P27
  - P59
  - PN71
  - PN58
  action: a-b-test
  since: '2026-06-19'
  notes: 'Cluster of OPEN qwen3 reasoning/tool-parser edge-case fixes found 2026-06-19 (loop scan), NOT
    vendored, candidates for the next pin bump: #38864 (</think> leak in streaming WITH stop sequences
    — token/text desync; distinct from our PN66 multiturn-leak), #45164 (Qwen3 parsing when reasoning
    end-tag present), #45787 (count reasoning tokens when <think> present — our P89 area), #44812 (stop
    thinking budget at implicit tool calls — our PN16 area), #45299 (reasoning->content short resp), #40861
    (Qwen3 XML/Coder streaming tool call). #39055 (tool calls embedded in think) + #45389 (braces in required-tool
    streaming) ALREADY referenced in registry. All OPEN, all edge cases. Live probe of #38864 on the 35B
    was INCONCLUSIVE (default traffic did not trigger reasoning_content + an SSE-capture issue) — no confirmed
    current exposure on the validated config (tool-call 7/7). Evaluate + vendor the applicable ones when
    a pin carrying them is pulled.'
- upstream: vllm#43730
  status: open
  local_patches:
  - P23
  - P87
  - G4_08
  action: watch
  since: '2026-06-19'
  notes: '[Bugfix] Marlin MoE c_tmp FP32-reduce IMA (drops a min() clamp in csrc/moe/marlin_moe_wna16/ops.cu;
    re-file of closed #36889). NOT overlay-patchable — the fix is in a COMPILED .cu kernel, not Python,
    so it can only arrive via a pin bump that carries it. Reproducer is RTX PRO 6000 Blackwell SM 12.0;
    our fleet is SM 8.6 A5000 and the live 35B runs c_tmp/marlin-IMA-clean. It is the WNA16 (INT4) MoE
    path (more the 27B than the 35B FP8). No confirmed current exposure; pull it when a pin includes it.
    Our P23 (marlin_fp32_reduce_wire) + P87 (sub-tile pad) + G4_08 (kdim pad) touch adjacent Marlin config
    but not this .cu clamp. 2026-06-19 loop MoE/quant scan.'
- upstream: vllm#45656
  status: merged
  local_patches:
  - P87
  - P91
  action: drift-check
  since: '2026-06-19'
  notes: '[Bugfix] Restore is_sym guard for zp in GPTQ/CT MoE — fixes a symmetric-quant regression
    introduced by #43409 (CPU W4A16 INT4 MoE, merged 2026-06-12, IN dev148). #45656 merged
    2026-06-18T16:20Z, AFTER our dev148 cutoff (b4c80ec0f, 04:18Z) -> next-pin candidate. The
    regression only mis-feeds qzeros to Marlin when is_sym=True (symmetric GPTQ MoE). Our 27B/35B
    are AutoRound INT4 ASYMMETRIC (real per-group zero-points present; P87 pads qzeros, P91 tags
    scales/qzeros — AutoRound emits these only for sym=False), so is_sym=False and the #45656 use_zp
    path is byte-identical pre/post fix -> NOT exposed. Drift-check only: when a pin carries #45656,
    re-confirm our checkpoints still load is_sym=False (no behavior change expected). 2026-06-19 loop
    MoE/quant scan.'
- upstream: vllm#45466
  status: merged
  local_patches:
  - P67
  action: watch
  since: '2026-06-19'
  notes: '[Bugfix][Kernel] Check output alignment in vectorize_with_alignment — fixes a misaligned
    store for non-16B-multiple head rows in reshape_and_cache_flash (fuzz model head_size=46). Merged
    2026-06-18T16:58Z, AFTER our dev148 cutoff -> next-pin. PR states "no behavior change for any
    currently-working caller". Our head dims (128/256/512) are all 16B multiples -> NOT exposed; P67
    next_power_of_2(hpk) handles the 27B GQA=6 grouping at the block level, separate from this per-store
    alignment. Track only so a future odd/non-pow2 head-dim model would not silently hit the misaligned
    store landmine. Free on bump. 2026-06-19 loop kernel scan.'
sweep:
- pr: 46384
  genesis_patch: PN346, PN346B
  trigger: reanchor-on-merge
  note: 'INCOMING ANCHOR-DRIFT (deep re-study 2026-06-26, dev424 3f5a1e173). #46384 ("[2/N]
    partial prefix-cache hits in hybrid coordinator") is an UNCONDITIONAL structural rewrite of
    find_longest_cache_hit in BOTH v1/core/kv_cache_coordinator.py AND
    v1/core/single_type_kv_cache_manager.py — the exact files+regions PN346 (manager half, vendor
    #43650) and PN346B (coordinator curr_hit_length clamp #45614 + Part-B FA/Mamba post-trim belt
    #46281) anchor into with required=True sub-patches. NOT a runtime bug for us today: its partial-hash
    DATA path is gated on enable_partial_hash_hits (needs mamba_cache_mode="align"; PROD runs "none" so
    the feature is inert). BUT the rewrite is mode-independent — it DELETES the _get_block_hashes closure,
    reshapes the clamp region (_new_hit_length -> get_cache_hit_length, cdiv truncation,
    KVCacheBlockListWithHitLength), and edits the MambaManager loop region. On the pin that carries #46384
    PN346/PN346B required anchors will be ABSENT -> the transactional patcher SKIPs -> silent LOSS of the
    #43559 Mamba+EAGLE/MTP+APC poison guard on the exact 35B-A3B MTP+APC path #43559 targets. On merge:
    re-anchor PN346 onto the new MambaManager partial branch; for PN346B check whether the merged clamp
    (get_cache_hit_length + cdiv) already encodes the monotonic-min — if so PN346B Part-A is a RETIRE
    candidate, else RE-ANCHOR onto the new clamp shape; re-derive the Part-B belt against the cdiv
    truncation. The anchor-SOT build_manifest already surfaces this mechanically (PN346/PN346B land in
    genuine_anchor_drift on the carrying pin); this row records the re-anchor-vs-retire DECISION + the
    #43559 guard rationale the mechanical detector cannot infer. Also tracked in the watch: section above.'
- pr: 46446
  genesis_patch: PN129
  trigger: retire-on-merge
  note: 'RETIRE-ON-MERGE (deep re-study 2026-06-26). #46446 ("Warm up slot-mapping kernel through
    BlockTable") adds BlockTable.warmup_compute_slot_mapping() (+ MultiGroupBlockTable fan-out) called
    from gpu_worker.compile_or_warm_up_model before cudagraph capture — the CLEAN upstream of PN129 (both
    vendor #42165). PN129 monkeypatches Worker.compile_or_warm_up_model and pokes the private Triton
    do_not_specialize attribute (self-documented hacky); #46446 does it the supported way (real
    compute_slot_mapping over multiple shapes via BlockTable, covering more specializations, no private-API
    poke) and runs for MultiGroupBlockTable which our hybrid (FA+GDN+Mamba) uses. PN129 is default-OFF and
    NOT enabled on prod YAMLs (no correctness exposure; boot-JIT latency only). On merge: retire PN129
    entirely. OPTIONAL before merge: re-base PN129 onto #46446 warmup_compute_slot_mapping (drop the
    private-Triton hack). Not urgent.'
- pr: 42006
  genesis_patch: G4_T1
  trigger: retire-on-merge
  note: 'G4_T1 RACING CLUSTER (#42006 / #42237 / #42300 / #44741 / #45068 / #44844) — whichever merges
    first decides retirement of the overlay stack (now THREE on-disk overlay generations: v1 rollback
    + v2 live mount + v3 prep). This PR = v1 overlay g4_t1_gemma4_tool_parser_pr42006_overlay.py (rollback
    copy). Also tracked in the legacy watch: section above.'
- pr: 42237
  genesis_patch: G4_T1
  trigger: retire-on-merge
  note: 'G4_T1 RACING CLUSTER member — v2 overlay (live bind-mount; 35/35 tool calls with Connection:close).
    Also tracked in the legacy watch: section above. On merge delete overlay files + launcher -v mount.'
- pr: 42300
  genesis_patch: G4_T1
  trigger: retire-on-merge
  note: G4_T1 RACING CLUSTER member — third competing upstream fix for the Gemma-4 streaming tool parser;
    not vendored; its merge also triggers overlay retirement review.
- pr: 44741
  genesis_patch: PN375
  trigger: retire-on-merge
  note: 'G4_T1 RACING CLUSTER member. DONE in the 2026-06-11 sweep: vendored as PN375 (pn375_gemma4_multiboundary_streaming.py)
    — _extract_streaming_delta_segments runtime hook rebinding _extract_streaming on the PRISTINE pin
    parser; the G4_14 pad-token set IS stripped from current_text AND delta_text before the consistency
    check (mutation-verified: without the strip the fix silently degrades to the single-pass path). Self-skips
    on the G4_T1 v2 overlay variant (accumulated-rescan, structurally immune). Combined regression test
    (multi-boundary + <pad> + G4_14 active) in tests/unit/integrations/tool_parsing/ test_pn375_gemma4_multiboundary_streaming.py;
    the pristine-bug reproduction test flips to FAILED when an upstream fix lands — then deep-diff and
    retire PN375 (iron-rule-#11).'
- pr: 45068
  genesis_patch: G4_T1
  trigger: retire-on-merge
  note: 'G4_T1 RACING CLUSTER member — Gemma4 parallel tool calls + MTP. DONE in the 2026-06-11 W2 sweep:
    (1) MTP-split regression corpus ported into test_g4_t1_streaming_mtp_corpus.py (first streaming parser
    tests for the family; runs against v2 AND the v3 prep), (2) token-id start-gate vendored into the
    v2 overlay (gate on tool_call_start_token_id in current_token_ids, Genesis adaptations documented
    inline), (3) keep-alive state leak (31/35 = remaining 4/35 failures on Gemma-4 + MTP) reproduced as
    a strict-xfail on v2 with the `if not previous_text` reset-guard analysis; the FIX ships only in the
    v3 prep overlay (Genesis reset-guard hardening) pending the server-stage A/B.'
- pr: 45207
  genesis_patch: G4_60E
  trigger: reanchor-on-merge
  detection: manual
  note: 'Mamba page padding — patches the SAME function (unify_kv_cache_spec_page_size) that G4_60E owns,
    with the same page_size_padded technique. W2 plan: fold the MambaSpec elif into G4_60E (standalone
    vendor would double-monkey-patch) + add Genesis-unique 3-way test (TQ + Mamba + bf16-drafter). On
    merge re-derive G4_60E anchors and drop the duplicated branch. detection=manual: G4_60E sub-patches
    are required=False (independent branches), so an incoming anchor drift soft-skips rather than failing
    genuine_anchor_drift — caught by the planned operator reconciliation pass over the shared function,
    not the anchor-SOT manifest.'
- pr: 45181
  genesis_patch: G4_60E
  trigger: reanchor-on-merge
  detection: manual
  note: 'Mixed KV page sizes (DFlash) — generic AttentionSpec padding fallback supersedes G4_60E Patch
    2 (keep TQ-native Patches 1/3/4); same function as #45207, do ONE reconciliation pass over G4_60E
    + the pr42637 overlay. Unblocks the Gemma-4-31B TQ k8v4 profile on this pin (~2.7x KV memory vs fp16).
    detection=manual: same shared-function reconciliation as #45207 (G4_60E sub-patches required=False).'
- pr: 45199
  genesis_patch: 'planned: PN371'
  trigger: retire-on-merge
  note: 'Deferred ref-pinned encoder-cache eviction — W1 vendor, flag-gated ON for gemma4 (engine-fatal
    "Encoder cache miss" on Gemma-4 vision + MTP K=3 + async-scheduling, our exact #38551 triple). Self-skip
    probe on eager_eviction signature; EXTEND: demote fatal assert to warn+skip in drafter path only.'
- pr: 45202
  genesis_patch: 'planned: P86'
  trigger: retire-on-merge
  note: Prefix-cache stats double-count on retries — W2 vendor as the safer variant (suppress recording
    in get_computed_blocks, record on allocate_slots success); fallback-disable if a connector is configured.
    Requires re-baselining p85/TQ hit-rate numbers collected under max_num_seqs=8.
- pr: 45182
  genesis_patch: watch-only
  trigger: review-on-merge
  note: TRTLLM BF16 MoE modular kernel is SM100-only (dead on Ampere) — borrow ONLY the logger.info_once("Using
    %s experts") one-liner into fp8/int_wna16 oracles + selection test asserting MarlinExperts (guards
    PN96b/PN368 against silent selection drift across pin bumps).
- pr: 45197
  genesis_patch: watch-only
  trigger: review-on-merge
  note: Disable cudagraph memory estimate default — upstream CHANGES_REQUESTED; do NOT flip configs blindly.
    MEASURE =1 vs =0 on 35B/27B first; if >~200 MiB overestimate confirmed, write the root-cause capture-all-and-measure
    patch (descs <= 16) and reconcile presets.py(=0) vs a5000 YAML(=1) inconsistency. Interacts with P66
    capture-size filter.
- pr: 45196
  genesis_patch: 'planned: PN-tbd (loader fail-fast validation)'
  trigger: retire-on-merge
  note: Loader/LoadConfig fail-fast validation — W2 near-clean 4-hunk vendor + tests; safety prerequisite
    for the enable_multithread_load + num_threads=8 experiment on 35B/27B (~30-60s saved per restart);
    mirror checks into audit_config_keys.py.
- pr: 45184
  genesis_patch: watch-only
  trigger: review-on-merge
  note: Intermediate tensors for KV connectors — W3 research. If vendored, FIX the upstream kwargs-shadowing
    bug first; the real win is a connector-free salience tap at the same decorator site (SnapKV-style
    scores as PN95 demote-ordering key; est. 5-15% fewer preemptions at high KV pressure).
- pr: 45176
  genesis_patch: watch-only
  trigger: review-on-merge
  note: 'Marlin -> torch stable ABI — byte-identical kernels, 0 direct perf. On the pin that contains
    it: verify _C_stable_libtorch/_moe_C op resolution before patch probes and re-run the Marlin A/B (all
    sm80 kernels recompiled). Adopt the pattern for genesis_kernels.abi3.so (native p87, fused TQ dequant,
    native pn365 GDN GEMM).'
- pr: 45173
  genesis_patch: watch-only
  trigger: review-on-merge
  note: '/v1/embeddings chat messages — Track A now: copy the ~20-line duck-typing into Genesis_proxy_ai
    to replace the zero-vector stub; Track B: vendor when the embedding sidecar lands (likely merged by
    next bump — review then). Unblocks semantic memory (MTEB ~64-70 via Qwen3-Embedding).'
- pr: 45100
  genesis_patch: PN370
  trigger: review-on-merge
  note: 'DONE 2026-07-05: the retire-on-merge intent EXECUTED — #45100 MERGED
    2026-06-22 and is native in both live pins; PN370 AND its 0.23.x twin PN398
    retired via pristine-dev748 deep-diff (see registry superseded_by). Row kept
    for provenance (PN400/#45656 precedent). Original entry:
    Racy accepted-counts under async+MTP — VENDORED 2026-06-11 as PN370 (spec_decode/pn370_async_accepted_counts_race.py):
    live silent-corruption class on Qwen3.6-35B PROD (hybrid GDN + MTP K=3 + async is the exact repro
    config; bonus ~2-5% TPOT from deleting a per-step synchronize). PN341 sub-patch 4 anchors the IDENTICAL
    num_accepted_tokens_event line — dual anchor variants (post-PN341 anchor imported from PN341_PREPARE_NEW,
    chain convention), PN341 dispatches first; PN290 composes cleanly. Drift markers self-skip on merge
    (needs_cpu_accepted_counts + the PR''s gdn_attn comment). Retire via deep-diff on the first pin containing
    the merge (iron rule 11).'
- pr: 45146
  genesis_patch: 'planned: P79e'
  trigger: retire-on-merge
  note: 'Reset placeholders on KV-load-failure rewind — W3, dormant until KV offloading is adopted (copy
    P79d TextPatcher skeleton; EXTEND with discard-credit for in-flight async frames). The study surfaced
    P79d staleness — W2-IMMEDIATE: rewrite P79d with discard-credit semantics (it backports a dead boolean
    and would assert-crash at async_scheduler.py:60 if enabled).'
- pr: 45144
  genesis_patch: P67b
  trigger: review-on-merge
  note: 'ROCm AITER SKV (MTP + fp8 KV) — P67b DRIFT-WATCH: second independent upstream validation of the
    P67/P67b multi-query design (PROD-active, +32% TPS); converge P67b toward its shape at next bump.
    Also the blueprint for Variant B (multi-query direct routing in the pr42637 TQ overlay to unblock
    MTP K=3 x TQ on Gemma-4-31B); its acceptance data (K=2: 0.75, K=3: 0.50) is a free K=2-vs-K=3 bench
    experiment.'
- pr: 45126
  genesis_patch: P81
  trigger: review-on-merge
  note: NVIDIA-tuned tiles + PID swizzle for triton_scaled_mm — probe whether the W8A8 fallback ever fires
    (likely dormant); main value is the transfer recipe (offline sweep -> frozen per-arch dict, bit-identical
    assert) for g4_kpad_moe_gemm. P81's table was tuned on GB10, NOT sm_86 — re-sweep; consume static
    tables via the PN362 force-first mechanism.
- pr: 45151
  genesis_patch: PN351
  trigger: reanchor-on-merge
  note: GUARANTEED PN351 ANCHOR-BREAKER — inserts 7 kwargs precisely inside the PN351_LAUNCH_OLD anchor.
    Make the PN351 anchor dual-form BEFORE any pin containing this merge (else we silently lose PN351's
    -3-7% decode TPOT). Also adopt the 3D-decode guardrail (forcing the 2D path = up-to-7x decode regression
    at our exact operating point); architecture is the GDN-epilogue-fusion template.
- pr: 45120
  genesis_patch: watch-only
  trigger: review-on-merge
  note: Fused softmax in grouped_topk CUDA — compiled .cu, cannot overlay; pin rejects scoring_func=2.
    Port the A/B method into a routing microbench at batch 1-8; the fp32-fused finding is retire evidence
    for kernels_legacy/router_softmax.py (dormant on the FusedTopKRouter path).
- pr: 45053
  genesis_patch: PN95
  trigger: reanchor-on-merge
  note: 'OffloadingHandler -> OffloadingWorker refactor renames symbols PN95 borrows — on post-#45053
    pins re-anchor the PN95 stream pool and adopt explicit submit_demote/submit_promote wrappers (replacing
    the string direction= kwarg anti-pattern upstream kills). With #45146, a prerequisite for any future
    CPU-KV-offloading experiment; track #33689''s hybrid-model item vs the PN95 MambaSpec exclusion.'
- pr: 45109
  genesis_patch: watch-only
  trigger: review-on-merge
  note: AWQ outputs under Transformers v5 tokenizer — test-only PR, vendor nothing. Build the tokenizer-fingerprint
    gate (sha256 of token-ids on a canonical prompt set, diffed pre-bench at every pin bump) BEFORE output
    benches that validate parser changes or PN351 anchor migrations — tokenizer drift otherwise gets misattributed
    to patches.
- pr: 45130
  genesis_patch: 'planned: PN-tbd (FP8 MoE + LoRA fail-fast guard)'
  trigger: retire-on-merge
  note: Fail-fast FP8 MoE + LoRA guard — 38-line additive, byte-verified clean on the pin, lands on our
    actual MarlinExperts class; vendor at next bump. Transfer the fail-fast pattern to TQ-k8v4+FA3 and
    Gemma-4 TQ-blocked drift guards; 5-min phantom-env audit (VLLM_MOE_FORCE_MARLIN is fake).
- pr: 45005
  genesis_patch: 'planned: PN-next (eagle_step zero-seqlen guard)'
  trigger: retire-on-merge
  note: 'W1 vendor: 8-line Triton guard (seq_len <= 0, stricter than upstream''s == 0) in the eagle_step_slot_mapping
    kernel; rebind in utils + llm_base_proposer. Kills #40756-class CUDA IMA on 262-280K MTP agent sessions;
    if A/B confirms, retires P108''s draft-loop sync workaround (2-6% TPOT recovery). PN id assigned at
    vendor time (PN370 collision flagged across chunks).'
- pr: 45060
  genesis_patch: 'planned: PN-next (all-NaN logits OOV kernel half)'
  trigger: retire-on-merge
  note: Vendor the KERNEL half only (tl.where -inf mask on the vocab padding tile); the scheduler hunk
    is the same site as PN133 — keep PN133 and extend it with log.error instead of upstream's assert.
    Qwen vocab 151936 % 8192 != 0, so the persistent-NaN livelock hole is live; ~0 latency cost.
- pr: 44993
  genesis_patch: P62
  trigger: review-on-merge
  note: 'Grammar advance across reasoning boundary — functionally COVERED by P62 (vendored #36138; its
    prefix scan is stronger for multi-token markers). Validation-only adaptation: port the 50-trial E2E
    json_object reproducer + 4 regression tests against P62; add drift markers for #44297/#44993 to protect
    P62 from a wedged apply at pin bump.'
- pr: 45040
  genesis_patch: 'planned: G4-family vendor (fp8_e5m2 KV predicate)'
  trigger: retire-on-merge
  note: 'Allow fp8_e5m2 KV for weight-only checkpoints — W1: monkeypatch the _checkpoint_has_fp8_kv_scales
    predicate into _init_kv_cache_quant, then a gemma4-31b-fp8e5m2-fallback profile (Triton-attn/FlashInfer,
    NOT FA2). 31B KV halves (~9.4 -> 4.7 GiB @200K) -> full 256K ctx or CG/concurrency headroom. Sequence
    with #45038: guard first, then fallback profile; check drafter KV-spec patches (G4_71/72/76) before
    enabling on structured role.'
- pr: 45038
  genesis_patch: G4_31
  trigger: retire-on-merge
  note: 'Guard fp8 KV auto-override on sub-SM90 — W1: extend G4_31''s Attention.__init__ wrap with a second
    suppress arm (cache_dtype=="auto" + kv_cache_scheme present) + late-mutation invariant log; do NOT
    double-wrap. Removes the IMA-on-burst landmine on the live 31B kv-auto degraded profile (SM8.6, MTP
    K=3, max_num_seqs=8). Audit G4_31 for redundancy on this pin (its turboquant_* arm can no longer trigger).'
- pr: 44955
  genesis_patch: 'planned: PN-next (parallel_tool_calls null guard)'
  trigger: retire-on-merge
  note: 'parallel_tool_calls null != false — W1: 1-line `is not False` text patch in tool_calls_utils.py
    modeled on PN288 (adjacent established territory); add the streaming-delta unit test upstream lacks;
    optional null-strip in Genesis_proxy_ai. Stops silent multi-tool-call truncation for explicit-null
    clients (LiteLLM/n8n). PN id assigned at vendor time.'
- pr: 45080
  genesis_patch: 'planned: PN-tbd (DecodeBenchConnector GDN fix)'
  trigger: review-on-merge
  note: 'DecodeBenchConnector list/tuple KV fix — our W2 backport DOES MORE than upstream (per-block fill:
    the pin''s MambaSpec IS block-indexed, upstream''s whole-pool fill clobbers concurrent state; plus
    a real group_idx map), so on merge deep-diff and KEEP the extras (iron-rule-#11 outcome b). Unlocks
    decode-TPOT-vs-depth benching on GDN hybrids (8K/32K/128K/280K sweep profile, MTP off).'
- pr: 45022
  genesis_patch: watch-only
  trigger: review-on-merge
  note: Voxtral realtime re-anchor RFC — all 3 gates fail on our stack; vendor nothing. Port vram_probe.py
    -> endurance_probe.py into the pin-validation playbook (catches multi-hour VRAM/RSS/KV creep on the
    24GB rig); keep the in-place-KV-mutation handshake as the design template for future live-KV-rewrite
    patches.
- pr: 45096
  genesis_patch: watch-only
  trigger: review-on-merge
  note: 'Mamba PD async-sched NIXL race — zero code now (NIXL dormant). Pin-bump checklist invariant:
    off-stream GDN state-pool writes MUST be event-gated (audit confirmed P7/PN204 side-streams already
    are); reuse the event + FIFO-deque + query() idiom when building GDN state prefix-cache restore (the
    4.4s@32K TTFT attack).'
- pr: 45001
  genesis_patch: watch-only
  trigger: review-on-merge
  note: 'CUDA qkv_padded_fp8_quant for ViT — dead on SM8.6 + text-only + PN62 skips ViT; do not vendor
    the .cu. Adapt the bandwidth playbook (BLOCK_M rows/program, grid-stride, vectorized loads) to the
    TQ k8v4 Triton kernels, gated on a <50%-of-768GB/s microbench; copy the bit-exact test matrix. #45080''s
    connector is the validation tool for any kernel rework derived from this.'
- pr: 44877
  genesis_patch: G4_T1
  trigger: retire-on-merge
  note: 'Quoted-key STRING_DELIM strip — DONE in the 2026-06-11 sweep: the ~25-line quoted-key branch
    transplanted verbatim into BOTH live-bindable G4_T1 overlay copies (v2 PR42237 current mount + v1
    PR42006 rollback; the sndr_private archive copy is frozen history, not live-bindable), superseding
    the #44717 post-hoc strip (provably dead once the branch is in). qwen3_xml audited for the same key/value
    asymmetry — PRESENT (unescaped key interpolation + quote capture in the <parameter=> preprocess regex)
    and fixed as PN374 (pn374_qwen3xml_quoted_keys.py). Tests: test_g4_t1_dict_key_sentinel_strip.py (now
    parametrized over all THREE overlay generations) + test_pn374_qwen3xml_quoted_keys.py (15). FOLDED
    into the v3 prep overlay (g4_t1_v3_gemma4_tool_parser_pr44844_overlay.py, 2026-06-11) per the synergy
    note — one overlay, one watchlist row each. Same G4_T1 file as the racing cluster.'
- pr: 44844
  genesis_patch: G4_T1
  trigger: retire-on-merge
  note: 'G4_T1 RACING CLUSTER member — span-based streaming parser rewrite. VENDORED 2026-06-11 (W2) as
    the v3 PREP overlay g4_t1_v3_gemma4_tool_parser_pr44844_overlay.py: verbatim PR head (sha a0c0ba69be)
    + Genesis reset-guard hardening (mutation-verified load-bearing) + the #44877 quoted-key hunk folded
    in per the synergy note. NOT bind-mounted anywhere — the A/B vs v2 on the 7x5 tool-call harness happens
    at the server stage before any mount switch. Closes the MTP K=3 streaming corruption v2 cannot handle
    (stripped call:name{...} bodies) AND the keep-alive 31/35 leak (strict-xfail on v2, PASS on v3, in
    test_g4_t1_streaming_mtp_corpus.py).'
- pr: 44943
  genesis_patch: PN380
  trigger: retire-on-merge
  note: 'Qwen3.5/3.6 MTP pre-fused expert loader — VENDORED 2026-06-11 (W2) as PN380 (pn380_qwen3_mtp_prefused_expert_loader.py):
    six sub-patches on qwen3_5_mtp.py load_weights — 3 vendored (pre-fused mapping + detection + params_dict
    KeyError guard, adapted to the pin''s STATIC mapping list; the PR head builds it in a loop, hence
    alt_ckpt_name is the safe drift marker) + 3 Genesis-original (P29-style draft-weight load-coverage
    guard: one log.error on any ckpt/param coverage gap — the engine''s strict check is disabled for quantized
    models, i.e. all PROD). Anchors disjoint from PN348 (same file; both co-apply orders tested). Companion
    accept-rate floor (0.55, WARN) landed in genesis_bench_suite.py. On merge: deep-diff (iron-rule-#11)
    — expect category (b), our coverage guard does MORE; keep the guard subs, retire the 3 vendored subs.'
- pr: 44837
  genesis_patch: watch-only
  trigger: review-on-merge
  note: 'DSV4 MTP missing prefix= — lesson ADOPTED 2026-06-11 (W2; not the code — the DSV4 files are outside
    our engaged set): AST lint tests/unit/lint/test_quantized_linear_prefix.py flags parallel-Linear-family
    constructions passing a non-None quant_config without prefix=, across (a) pristine pin model files
    our patches anchor (auto-discovered from patch sources) and (b) vendored overlay copies in the repo;
    scoped allowlist for verified-safe hits (EMPTY as of pin g303916e93 — whole surface clean). Kills
    one pin-bump crash class + the silent AWQ/AutoRound mis-quantization variant; <1s CI. Built together
    with #44943''s coverage guard ("loud startup" validation family, PN380).'
- pr: 44880
  genesis_patch: PN340, PN341
  trigger: reanchor-on-merge
  detection: test
  note: 'Bailing MTP + LINEAR_ATTN spec-decode — do NOT vendor the feature (the pin''s gdn_attn already
    has the design). detection=test: PN340/PN341 sub-patches are required=False (independent), so an
    incoming anchor drift soft-skips rather than failing genuine_anchor_drift — caught instead by the
    PN370 chain drift tests (tests/unit/integrations/spec_decode/test_pn370_async_accepted_counts_race.py)
    + the PN340 buffer-stability test, which import PN341/PN340 anchor constants and fail loudly on drift.
    ANCHOR BREAKER: the PR edits PN341''s anchor neighborhood in gpu_model_runner.py
    (+10/-2 around the accepted-counts / spec-decode preprocessing region PN341 sub-patch 4 and PN370
    both anchor) — GUARANTEED PN341 anchor break on the merge pin; this row converts it into a planned
    5-min re-anchor (re-derive PN341_PREPARE_OLD/NEW, then re-run the PN370 chain tests, which import
    PN341''s constants and fail loudly on drift). Buffer-stability tests PORTED 2026-06-11 (W2 sweep):
    tests/unit/integrations/attention/gdn/ test_pn340_gdn_builder_buffer_stability_torch.py — data_ptr
    identity across consecutive FULL-CG builds + pad semantics (upstream''s PAD_SLOT_ID adapted to the
    GDN builder''s NULL_BLOCK_ID / False / 1 / last-cumsum constants) against the PN340-patched builder,
    K=1 upstream shape + K=3 PROD shape. REMAINING at next pin bump: cherry-pick the PR''s llm_base_proposer.py
    getattr-hardening (defensive getattr(..., None) reads of optional drafter attrs — cheap crash-class
    insurance for non-Bailing MTP proposers too).'
- pr: 44868
  genesis_patch: 'planned: PN358'
  trigger: retire-on-merge
  note: 'FULL CG forward-context refresh — W2 vendor as PN358 (journal pre-assigned) with 2 improvements:
    data_ptr-pruned copy (kills upstream''s 1-3% TPOT cost) + a `detect` mode logging stale-metadata mismatches
    (definitive audit of whether the 280-patch overlay leaks fresh tensors into captured graphs). Composes
    with PN353B/PN118, no file conflict.'
- pr: 44850
  genesis_patch: 'planned: PN-tbd (tile_mask USE_TD KV load)'
  trigger: retire-on-merge
  note: tile_mask in USE_TD KV load — W3 vendor of the 4-line tl.where fix (constexpr-dead when USE_TD=False,
    provably free); shares a file with PN351 but anchors verified non-colliding. Adopt the NaN-canary
    harness and run it against tq_grouped_decode/sparse_v/p67 kernels; makes VLLM_TRITON_ATTN_USE_TD A/B
    safe.
- pr: 44912
  genesis_patch: PN77
  trigger: review-on-merge
  note: '0-D FP8 scale rank vs torch._scaled_mm — upstream kernel dormant on SM8.6 (Marlin tier selected);
    the SAME bug class lives in OUR lm_head_fp8_method.py:416-434 (amax() produces 0-D scales): fix the
    Genesis side (.view(1) + TDD unit test) + grep-audit CI for all _scaled_mm call sites (inoculates
    pn347/p81/compressor paths).'
- pr: 44784
  genesis_patch: 'planned: PN-tbd (KV-offload MTP draft-group gate)'
  trigger: retire-on-merge
  note: 'KV-offload + MTP segfault — OffloadingConnectorScheduler wrongly schedules EAGLE/MTP draft-attention
    groups -> OOB GPU block index -> cuMemcpyBatchAsync segfault; blocks native KV offload on every MTP
    config (ours included). W1 doc half DONE 2026-06-11 (docs/PATCHES.md stale GDN-offload caveat corrected
    against the pin). W2: vendor the 4 scheduler.py hunks tests-first + EXTEND with Qwen3.6-specific is_eagle_group
    flagging (avoid all-groups fallback hit-rate loss) + re-add the dropped pre-DMA bounds check. Unlocks
    prefix-reuse (second-pass TTFT ~4-8x cut on agent loops).'
- pr: 44932
  genesis_patch: PN22, PN23
  trigger: review-on-merge
  note: 'FP8 KV + DCP FlashInfer — partial overlap with our PN22/23 dequant_buffer: borrow the fused gather+dequant
    pattern, keep the TQ specifics. Part 1 (S, anytime): cherry-pick the selector.py info_once log-dedup
    hunk. Part 2: TQ-k8v4 fused gather+dequant Triton kernel + narrow the PN22/23 dequant buffer dtype
    (fp32 -> fp16 V). Part 3 research: quantize-before-collective for PCIe TP=2. Bench together with #44784
    on the 35B long-ctx profile.'
- pr: 44717
  genesis_patch: G4_T1
  trigger: retire-on-merge
  note: 'Dict-key sentinel strip — DONE in the 2026-06-11 sweep: 8-line key-strip vendored into both G4_T1
    overlay copies (v2 PR42237 + v1 rollback), 10 new tests, rides the G4_T1 registry entry. SUPERSEDED
    same sweep by the #44877 quoted-key branch (same issue #44715, strictly stronger: handles '':'' inside
    quoted keys + withholds unterminated keys) — the overlays now carry #44877''s hunk; this row stays
    because EITHER PR merging still triggers the same retire review. #44752 is a DUPLICATE of this PR.'
- pr: 44752
  genesis_patch: G4_T1
  trigger: retire-on-merge
  note: 'DUPLICATE of #44717 (same bug, same issue #44715) — do NOT vendor separately (avoids a redundant
    PN); tracked so that EITHER PR merging triggers retirement of the G4_T1 key-strip hunk.'
- pr: 45076
  genesis_patch: PN367
  trigger: retire-on-merge
  note: 'Negative cudagraph memory estimate — DONE this sweep as PN367 v2 (never-firing drift markers
    fixed + 1 MiB first-capture floor); registry re-pointed to the consolidated #44745. Watch BOTH #45076
    and #44745; retire PN367 when either merges (deep-diff first). Prevents the capture/load OOM class
    under MTP K=3 at util=0.9 on 24GB cards.'
- pr: 44644
  genesis_patch: PN348
  trigger: retire-on-merge
  note: 'Qwen3.5/3.6 MTP backbone dedup — already vendored as PN348; ENABLED on the qwen3.6-35b-a3b-fp8
    profile 2026-06-11 (enable+measure: ~1 GiB/rank peak VRAM + 1-3s boot expected; A/B at next restart).
    Gain claim reframed per the sweep study: the pin''s proposer already reclaims steady-state — the benefit
    is PEAK load VRAM + boot time, steady-state TPS unchanged. With PN367, shrinks boot-time VRAM uncertainty
    on the same 24GB MTP boot path.'
- pr: 44742
  genesis_patch: PN381
  trigger: retire-on-merge
  note: 'allowed_token_ids metadata hardening — VENDORED 2026-06-11 (W2 sweep) as PN381 (anchored 1-line
    condition in _make_sampling_metadata, PN67 playbook; default OFF via GENESIS_ENABLE_PN381_ALLOWED_TOKEN_IDS_METADATA);
    defense-in-depth only (the #35654 consumer fix is already in the pin). On merge the patcher SELF-SKIPS:
    drift marker is the PR''s unparenthesized form ''or not self.no_allowed_token_ids'' — our emitted
    clause is parenthesized on purpose (lint_drift_markers contract); then deep-diff + retire per iron-rule-#11.
    The PR''s CUDA regression test is PORTED CPU-runnable through OUR PN369/P71-patched sampler (tests/unit/integrations/spec_decode/
    test_pn381_sampler_regression_torch.py) + a torch-free producer contract exec test — row-parity coverage
    upstream never had.'
- pr: 44778
  genesis_patch: PN55
  trigger: retire-on-merge
  note: 'FP8 KV wake-up nested containers — functionally DUPLICATE of existing PN55v2 (drift marker matches
    the helper); no re-vendor. Do: add #44778 to PN55 related_upstream_prs, replace the hand-mirrored
    test with the PR''s exec-patched-text technique, add log.warning for skipped non-tensor leaves, review
    companion #44779 (resume_scheduler gate) before enabling sleep/wake hot-swap. On merge deep-diff PN55v2
    for retirement (iron-rule-#11).'
- pr: 44628
  genesis_patch: 'planned: PN-tbd (fp8 ignore-list substring match)'
  trigger: retire-on-merge
  note: FP8 modules_to_not_convert substring match — W2 vendor (text-patch fp8.py at both call sites +
    quant_utils experts branch + parity one-liners); validate via per-layer quant-scheme log diff on 35B
    before default_on. Eliminates a silent-gibberish class for HF-style ignore lists (Qwen3.6-VL FP8 broken
    TODAY on the pin); unblocks selective-FP8 GDN re-quant experiments.
- pr: 44563
  genesis_patch: 'planned: PN-tbd (moe_wna16 BLOCK_SIZE_K clamp)'
  trigger: retire-on-merge
  note: 'moe_wna16 BLOCK_SIZE_K clamp for gs=32 — W2 vendor of the 4-line clamp (anchor unique, beside
    the P24 surface) + the PR''s CPU-only test + a boot-time legality assert for the actual model grid.
    Converts a deterministic warmup abort into serving for gs=32 int4 MoE on the live Marlin->wna16 fallback;
    unblocks gs=32 quant benchmarking of our exact model family. The #40547 heuristic-swap finding feeds
    the P24 MoE-tiling research track.'
- pr: 44754
  genesis_patch: watch-only
  trigger: review-on-merge
  note: 'Revert of #44613 snapshot — DO NOT vendor (misattributed auto-revert, abandoned, conflicts; #44613
    is correctly in the pin). Do: adopt the snapshot-at-construction discipline lint (rule for #44628
    and any FusedMoEConfig-adjacent patch: P24/PN352/PN96B), harden the bare get_current_vllm_config()
    at turboquant_attn.py:496 (same boot/reload AssertionError class dev134 already triggered once).'
- pr: 45413
  genesis_patch: P64
  trigger: review-on-merge
  note: 'IRON-RULE-#11 DEEP-DIFF PREP (added 2026-06-28). #45413 ("declarative state-machine streaming
    tool-call parser for Qwen3") MERGED 2026-06-15; Gemma4 follow-up #45588. This is a ground-up REWRITE
    of the Qwen3 streaming tool-call parser as a declarative state machine — it directly OVERLAPS the
    Genesis qwen3coder tool-parser defense stack: P64 (MTP streaming early-return fix, vllm#39598),
    P61c (deferred-commit-until-<function= header, club-3090#72) and PN56 (XML parse fallback, vllm#41466),
    all consolidated in sndr/engines/vllm/patches/tool_parsing/p64_p61c_pn56_qwen3coder_consolidated.py
    (genesis_patch=P64 is the live consolidated id; P61c/PN56 ride as env_flag_aliases on it). The shared
    failure mode is club-3090 #178 (MTP x qwen3_coder argument corruption) — a streaming tool-call ARG gets
    garbled when MTP speculative tokens cross a tool-call boundary; both the upstream state machine and our
    three patches exist to keep that boundary clean. trigger=review-on-merge (NOT retire-on-merge): per
    iron-rule #11 do NOT retire on a title/abstract match. ON THE NEXT PIN BUMP that carries #45413, deep-diff
    the merged state machine against P64/P61c/PN56 LINE BY LINE and classify each: RETIRE (the state machine
    subsumes the patch verbatim), PARTIAL (borrow the upstream shape, keep the Genesis-unique guard), or KEEP
    (the patch covers a case the rewrite does not). What to compare, concretely: (1) does the new state
    machine emit </function> exactly once per call without the P64 early-return on the MTP multi-token delta?
    (2) does it defer the commit until the <function= header is seen, or can it still leak a partial header
    like P61c guards against? (3) does it carry an XML parse-failure fallback equivalent to PN56, or will a
    malformed Coder XML block now hard-fail? (4) regression-gate the merged parser against our streaming MTP
    corpus (the #45068/#178 MTP-split tests) BEFORE retiring anything. Prep note with the full comparison
    checklist + rationale: docs/PR45413_QWEN3_PARSER_DEEPDIFF_PREP.md.
    review-on-merge is advisory and intentionally NOT registry-bound, so this prep row records the decision
    inputs the mechanical drift detector cannot infer; it does NOT pre-commit any retire/keep outcome.'
- pr: 42890
  genesis_patch: G4_60E
  trigger: review-on-merge
  note: 'NEXT-BUMP BOOT-BLOCKER SOURCE (deep-studied 2026-07-05; MERGED 2026-07-04T02:29Z — AFTER the
    dev748 cut 2026-07-03T04:11Z, so NOT in dev748; verified by pristine-tree grep on the rig:
    attn_utils.py has no KVQuantMode.NONE conditional, line 308 passes cache_dtype_str=cache_dtype
    straight through). #42890 (kv_cache_dtype_skip_layers) makes the v1 reshape pass
    cache_dtype_str="auto" whenever kv_quant_mode == KVQuantMode.NONE; TurboQuant dtypes map to
    KVQuantMode.NONE while TQFullAttentionSpec still needs the real turboquant_k8v4 string for the
    backend shape -> engine startup ValueError "Unknown TurboQuant cache dtype: auto" = HARD BOOT FAIL
    on --kv-cache-dtype turboquant_k8v4. BOTH heavy lanes run TQ k8v4 (27B INT4 TQ k8v4+PN520 ~130 t/s;
    35B where TQ k8v4 = +11.3% TPS), so ANY pin cut after 2026-07-04 is a boot-blocker unless the fix
    #47609 (see its own row) is also in the pin. Mechanical detection: UPSTREAM_MARKERS
    PR_42890_kv_skip_layers_cache_dtype_auto fires newly_merged on the carrying pin. SECONDARY impact
    bound to G4_60E: sndr/engines/vllm/patches/attention/turboquant/g4_60e_kv_cache_utils.py (vendor of
    OPEN vllm#45181) runtime-wraps vllm.v1.worker.gpu.attn_utils._reshape_kv_cache and carries a vendored
    MIRROR of its logic ("Mirror gpu/attn_utils._reshape_kv_cache") — when #42890 (+#47609) enters the
    pin that mirror is STALE and G4_60E must be re-studied (gemma4 padded-page lanes). No Genesis patch
    text-anchors attn_utils.py, so no anchor drift; the wrapped function SEMANTICS change.'
- pr: 47609
  genesis_patch: 'planned: TQ cache-dtype preservation backport (attention.turboquant) — only if still
    unmerged at bump time'
  trigger: review-on-merge
  note: 'PIN-BUMP PREFLIGHT GATE ITEM (deep-studied 2026-07-05; OPEN, author LucasWilkinson = maintainer,
    labels bug/ready -> likely merges). Fix for the #42890 TQ boot-blocker (see the #42890 row): adds
    "and not isinstance(kv_cache_spec, TQFullAttentionSpec)" to the two KVQuantMode.NONE conditionals in
    vllm/v1/worker/gpu/attn_utils.py (_reshape_kv_cache + _update_hybrid_attention_layout), 6 lines.
    NOT needed on dev748 (the #42890 region does not exist there — verified on the pristine rig tree).
    AT BUMP TIME, for any candidate pin cut after 2026-07-04: (1) gh pr view 47609 — if MERGED and in the
    candidate, nothing to do beyond the G4_60E mirror re-study; (2) if still UNMERGED, the candidate
    HARD-FAILS BOOT on turboquant_k8v4 lanes — backport as a NEW attention.turboquant patch anchored on
    the two "layer_cache_dtype = (" sites (BOTH sites needed; anchors structural; dual-variant NOT
    required since dev748 lacks the region entirely — the patch is next-pin-only with a lower
    vllm_version_range bound). Mechanical detection: UPSTREAM_MARKERS PR_47609_tq_cache_dtype_preserved
    fires newly_merged when the fix lands in a candidate tree.'
- pr: 47611
  genesis_patch: watch-only
  trigger: review-on-merge
  note: 'IRRELEVANT TO GENESIS (deep-studied 2026-07-05; OPEN). Handles older FlashInfer wheels lacking
    the gemm1_alpha/gemm1_beta/gemm1_clamp_limit kwargs that #45723 (MERGED 2026-07-01, IS in dev748 —
    verified unconditional gemm1_alpha= at trtllm_fp8_moe.py:234/405 in the pristine tree) passes
    unconditionally -> TypeError at startup on old FlashInfer (nightly CI: gpt-oss-20b 0.0 accuracy,
    Qwen3-30B FP8 startup timeouts). Fix = functools.cache-d inspect.signature probe; narrower
    alternative to the wholesale draft revert #47431 (OPEN). Dormant on our rig despite the regression
    source being in dev748: the TrtLlmFp8Experts dispatch is gated is_device_capability_family(100) +
    has_flashinfer_trtllm_fused_moe() (trtllm_fp8_moe.py:98-99) = Blackwell SM100 only; 2x A5000 = SM86
    never selects it. Our FP8 35B alt lane runs FP8 Marlin MoE (MarlinExpertsBase, PN352B registry
    evidence); 35B PROD is AWQ Marlin; no Genesis patch touches trtllm_fp8_moe.py (grep clean). No
    backport, no watch burden unless Blackwell hardware arrives.'
- pr: 47593
  genesis_patch: P24
  trigger: review-on-merge
  note: 'IRRELEVANT TO CURRENT STACK, ANCHOR-AUDITED (deep-studied 2026-07-05; OPEN). Changes
    get_default_config fp8_w8a8 block-quant [128,128] BLOCK_SIZE_M rule "M<=64" -> "M<=64 or
    M*topk<16*E" (keep the 16-row tile while the avg expert sees <1 full tile), CUDA-only; ROCm +
    batched per-expert-M callers keep the old rule via a new m_is_per_expert plumb. Benched geomean
    1.295x on H20 / 1.015x on 4090D at M=96-768 for high-E MoEs. NOT in dev748 (fused_moe.py:1285 still
    has the raw rule). The rewritten branch NEVER EXECUTES on our rig: Ampere FP8 MoE dispatches to FP8
    Marlin (not the triton fp8_w8a8 block branch); 35B PROD is AWQ/wna16-Marlin; G4 26B WNA16 MoE uses
    the int4_w4a16 branch (untouched, already keys on M*topk via should_moe_wna16_use_cuda); decode M at
    max_num_seqs=2 sits under the unchanged M<=64 floor anyway. ANCHOR COLLISION AUDIT: P24 anchor-1
    (_OLD_FP8_CFG) targets the EXACT line #47593 rewrites, but that anchor is ALREADY dead on dev748
    (pre-existing #46642-era refactor drift: "BLOCK_SIZE_N: block_shape[0]" -> block_n; anchor is
    required=False so it soft-skips today); only P24 anchor-2 (general branch) applies and #47593 does
    NOT touch that region. PN377 (_ensure_block_size_k_divisible, wna16 branch) disjoint. Net: if
    #47593 merges, ZERO additional drift. HOUSEKEEPING (pre-existing): P24 fp8-branch num_warps overlay
    is a silent no-op on dev748 — re-anchor only if the fp8 triton MoE branch ever becomes live for us
    (on Marlin lanes it is dormant by design, P24 docstring). Idea-family overlap with P81 (backport of
    OPEN vllm#40925 — same small-M-tile-for-decode reasoning but for the DENSE
    w8a8_triton_block_scaled_mm kernel, different file) = outcome (c) different problem, both keep.'
# ── 2026-07-05 batch-triage 47382..47564 (34 PRs: 4 backport-now PN523-PN526,
# 4 next-bump-gate, 4 anchor-collision, 22 irrelevant + 5 hygiene rows) ──
- pr: 47507
  genesis_patch: 'G4_04, G4_06, G4_07, G4_08'
  trigger: review-on-merge
  note: 'NEXT-BUMP GATE (deep-studied 2026-07-05; OPEN). Gemma4 k_eq_v x compressed-tensors shard-alias
    propagation: should_ignore_layer rewrite + SupportsQuant apply_checkpoint_shard_aliases MRO hook +
    extend_with_shard_aliases in compressed_tensors/utils.py. At the first pin containing it: boot-verify
    all 3 G4 CT lanes (quant method selection unchanged, apply failed=0) + preflight the gemma4 family
    G4_04/G4_06/G4_07/G4_08 against the new SupportsQuant MRO / configure_quant_config early-return.
    ESCALATE to backport if a partially-quantized G4 checkpoint (per-shard ignore entries) is adopted
    before merge — vendor keeping upstream model-agnostic get_checkpoint_shard_aliases shape, never
    hardcode k_eq_v indices. Mechanical detection: UPSTREAM_MARKERS PR_47507_gemma4_ct_shard_aliases.'
- pr: 47396
  genesis_patch: PN520
  trigger: review-on-merge
  note: 'NEXT-BUMP GATE (deep-studied 2026-07-05; OPEN). Per-shard per-tensor FP8 scales for fused GDN
    in_proj — linear.py weight_loader_v2 tuple branch (numel()==len(shard_id) probe). Fail-loud bug
    (parameter.py:305 assert) proven NOT tripped by our checkpoints (fleet rerun 2026-07-05 failed=0).
    At the merge pin: re-verify PN520 proof-of-life boot log ("[PN520] imperative load_weights ACTIVE" +
    shard routing counts) since its stacked_params_mapping exercises the changed branch; fix strictly
    beneficial to PN520, nothing of ours to retire. ESCALATE to backport-now ONLY if an FP8 GDN
    checkpoint shipping fused [q,k,v] per-shard weight_scale vectors is onboarded on a pre-fix pin
    (dual-variant TextPatch on the tuple branch, keep the upstream probe verbatim). Mechanical
    detection: UPSTREAM_MARKERS PR_47396_fp8_gdn_per_shard_scales.'
- pr: 47391
  genesis_patch: 'PN128, PN90, P108, PN357, PN372'
  trigger: review-on-merge
  note: 'NEXT-BUMP GATE (deep-studied 2026-07-05; OPEN). DELETES eagle_prepare_inputs_padded_kernel and
    rewrites padded EAGLE input prep as torch ops in llm_base_proposer.py — our live 35B MTP K=5 hot
    path. At the merge pin: (1) retire PN128 kernel-2 warmup arm (kernel gone; torch ops need no JIT
    warmup) with the deletion kept self-documented via UPSTREAM_MARKERS PR_47391_padded_eagle_torch_ops;
    (2) re-verify PN90/P108/PN357 anchors in llm_base_proposer + PN372 in spec_decode/utils.py (regions
    byte-checked disjoint, expect clean); (3) A/B decode_TPOT on 35B MTP K=5 per canonical bench since
    per-step prep is rewritten. Early vendoring LOW value — PN128 already killed the JIT spike.'
- pr: 47382
  genesis_patch: watch-only
  trigger: review-on-merge
  note: 'NEXT-BUMP semantics note (deep-studied 2026-07-05; OPEN). TPOT/ITL metrics mis-accounting under
    MTP multi-token steps (stats.py:400 records one sample per step). NO Genesis consumer reads
    vllm:inter_token_latency_seconds (bench discipline is client-side decode_TPOT; Prometheus scrape
    reads only spec_decode accept counters) -> do NOT vendor (carrying-cost calculus, iron-rule-#10
    exception). At the merge pin: ITL histogram semantics change — label any first server-ITL vs client
    decode_TPOT comparison with (pin, date). DONE NOW (this commit): stale SLO row
    "vllm:inter_token_latency_seconds < 10ms" in Genesis/docs/reports/GENESIS_LLM_DEPLOYMENT_REPORT.md
    flagged as not meaningful under MTP until the fix lands; stale PN361 "draft_sample_method=
    probabilistic" comment in the a5000-2x hardware YAML fixed (PROD draft is GREEDY — PN90
    probabilistic is version-gated OFF, PN390 credit correction 2026-06-16).'
- pr: 47556
  genesis_patch: watch-only
  trigger: review-on-merge
  note: 'HYGIENE (triaged 2026-07-05; OPEN; irrelevant verdict). Sliding-window prefix-cache miss-scan
    optimization. Only Genesis exposure: the pr42637 STALE CORE SHADOW under
    attention/turboquant/overlays/pr42637/ drifts further from upstream on merge — fatal ONLY if that
    overlay is ever re-mounted. No live lane mounts it; no anchor overlap. Re-read on merge, no action.'
- pr: 47459
  genesis_patch: watch-only
  trigger: review-on-merge
  note: 'HYGIENE (triaged 2026-07-05; OPEN; irrelevant verdict). Gemma4 audio crash with variable-length
    batched inputs. No G4 lane has audio input enabled -> unreachable. BECOMES backport-now the day
    audio input is enabled on any G4 lane on a pre-fix pin; re-check this row in that decision.'
- pr: 47460
  genesis_patch: watch-only
  trigger: review-on-merge
  note: 'HYGIENE (triaged 2026-07-05; OPEN; irrelevant verdict). Initializes draft CUDA-graph keys for
    the native draft_model proposer. We run MTP (no draft_model lane configured anywhere in the fleet).
    2-line gate extension IF a draft_model lane is ever configured on a pre-fix pin.'
- pr: 47386
  genesis_patch: watch-only
  trigger: review-on-merge
  note: 'HYGIENE (triaged 2026-07-05; OPEN; irrelevant verdict — probabilistic path OFF on our lanes).
    Decouples draft Gumbel stream from acceptance/recovery noise on the V2 runner. RULE: any
    probabilistic-draft re-evaluation on a V2 pin REQUIRES this fix (or its salt) first — otherwise the
    correlated-noise bias contaminates the A/B. The stale PN361 "our PROD draft_sample_method=
    probabilistic" comment in the a5000-2x hardware YAML is fixed in this commit (PROD draft is GREEDY;
    PN90 probabilistic version-gated OFF).'
- pr: 47445
  genesis_patch: 'PN376, PN520'
  trigger: review-on-merge
  note: 'HYGIENE (triaged 2026-07-05; OPEN; irrelevant verdict — no ModelOpt checkpoints in the fleet).
    ModelOpt quantization inference fix for fused siblings. On merge: re-run PN376 anchor preflight
    (fp8 modules_to_not_convert substring region adjacency) and add a line to the PN520 upstream-state
    block noting the fused-sibling loader semantics changed upstream.'
- pr: 47491
  genesis_patch: PN346B
  trigger: reanchor-on-merge
  note: 'ANCHOR COLLISION (deep-studied 2026-07-05; OPEN). Inserts 5 lines INSIDE the 4-line
    PN346B_ANCHOR_OLD span in kv_cache_coordinator.find_longest_cache_hit (between
    eagle_verified.clear() and the naked curr_hit_length assignment; byte_offset 28254 on dev748).
    PRE-ARMED (batch-triage STEP 0b): the PR''s two inserted comment lines are now PN346B
    upstream_drift_markers, so the merge classifies LOUDLY instead of as unexplained drift. ON MERGE —
    SEMANTIC reconciliation, not a naive re-anchor: (1) PN346B''s min() clamp must sit AFTER upstream''s
    new `continue`; (2) upstream''s attention-hit preservation SKIPS the hit_blocks_by_group update for
    the missing Mamba group -> a resumed request would skip N tokens with NO Mamba state (the
    boots-clean-but-garbage class) — suppress on our lanes; P85''s shadow fine-hash stays OUR mechanism
    (iron-rule-10 outcome: different-broken-approach-same-goal, documented).'
- pr: 47489
  genesis_patch: P62
  trigger: review-on-merge
  note: 'ANCHOR COLLISION (deep-studied 2026-07-05; OPEN). Rewrites the grammar.accept_tokens block that
    IS P62 sub-patch 2a''s SCHED_UPDATE_FROM_OUTPUT_OLD (byte-for-byte); PR context shows upstream main
    already carries the post-dev748 #44297 advance_token_ids refactor in the same region. Bug itself
    UNREACHABLE for us (zero min_tokens usage repo-wide) -> review-on-merge, not reanchor (P62''s
    existing #44993 watch markers convert the mismatch to a clean upstream-drift SKIP meanwhile). ON the
    bump past #44297/#44993/(47489 if merged): re-derive P62 sub-2a against the new block shape and FOLD
    grammar.is_terminated() -> FINISHED_STOPPED early-stop INTO P62''s replacement text — a naive
    re-anchor silently REVERTS the 47489 fix.'
- pr: 47513
  genesis_patch: 'P62, P70, PN275'
  trigger: review-on-merge
  note: 'ANCHOR COLLISION + STUDY ITEM (deep-studied 2026-07-05; OPEN). Touches scheduler.py
    update_from_output (P62 2a / P58 / P79C / P79D / PN58 family), config/speculative.py
    (p70_auto_strict_ngram, PN275) and the config/vllm.py dynamic-SD cudagraph override. At the first
    pin containing it: run preflight + re-verify those anchors. STUDY (iron rule #10 — borrow design,
    not code): its num_speculative_tokens_per_batch_size cost-table + bucket-constrained per-request
    width allocation is direct prior art for g_dynamic_k_mtp_proposer.'
- pr: 47462
  genesis_patch: 'G4_26'
  trigger: review-on-merge
  note: 'COLLISION TERRITORY, G4_26 retired (deep-studied 2026-07-05; OPEN mega-PR, 4777-line diff — do
    NOT vendor). DiffusionGemma custom_sampler call site + logits_processor + V2 runner rewrite with
    VLLM_DIFFUSION_GEMMA_LOCAL_VOCAB_SAMPLER. STEP 0a already executed the iron-rule-11 audit this PR
    surfaced (contradiction C1): pristine dev748/dev714/dev672 REPLACED G4_26''s anchor site with the
    native #46177 local-shard soft-embed + all_reduce -> G4_26 retired (superseded_by #46177), flag
    removed from the diffusiongemma ModelDef, profile verify-note fixed. ON 47462 MERGE: evaluate
    VLLM_DIFFUSION_GEMMA_LOCAL_VOCAB_SAMPLER=1 behind the canonical bench gate ONLY after a
    block-diffusion-aware harness exists; keep native csrc ops OFF (no SM86 exposure).'
__sentinel__: complete