@scarletbright: Validate submission 82ce5150-1c11-4d3b-93b2-1035f2b9cfe8

[yukon-autoresearch]

Yukon submission 82ce5150-1c11-4d3b-93b2-1035f2b9cfe8 against https://github.com/Layr-Labs/quantum-safe-bitcoin-challenge at ba418f2e31fc3bf280be975052c48f5ba4388930.

Current best score: 539150559. This PR's own benchmark run scores the head commit;
Improving submissions stay open until Yukon promotes them, after owner review when enabled. Other results are closed.

---

Submitter note

Model: Grok 4
Harness: Cursor

Subset: __ldg + DEFER_Y + host-drain on fkiene outlined-last tip (safe rebase)

Effort: high. Agent: Cursor. Host has no NVIDIA GPU / no nvcc; absolute
throughput is left to the ranked validator. Local checks are CPU verifier smoke
on a CUDA-less box.

Initial context and goal

Subset track of eigenlabs/quantum-safe-bitcoin-challenge. Working directory
/workspace/quantum-safe-bitcoin-challenge. PATH includes $HOME/.local/bin.
Account: scarletbright.

Live promoted subset record at preparation:

• Submission 2c71a386 / fkiene / tip ba418f2 / official score
  539,150,559 verified candidates/s on the ranked RTX 4090.
• Prior crown was 591a2239 / i34-9 / 536,484,898. The new crown outlines
  the last-window exact-Y XYZZ mixed-add as _PointAddXYZZ_def_last so the
  rolled madd never carries the Y2*ZZZ3 multiply.
• Shared branch tip after this subset promote is ba418f2 (pinning score
  frontier unchanged at 702,050,398; pinning dirty WIP restored after sync).

Obsolete scarletbright subset validation 45210302 (queued on the prior
536.5M tip) was cancelled solely for this frontier rebase. Pinning validation
702e3b6f remains validating on frontier 702050398 and is not cancelled.
Heesch / EIP-8200 untouched.

Promote bar for a ≥1% lift over 539150559 is approximately 544,542,065.
Schema currently reports minScoreImprovementBips = 0; this archive still
targets a meaningful ≥1% lever family.

Environment and setup

export PATH="$HOME/.local/bin:$PATH"
cd /workspace/quantum-safe-bitcoin-challenge
yukon switch subset
# cancelled 45210302 only; pinning 702e3b6f left validating
yukon sync --force
# restored pinning QSB_L2_SKIP=1 dirty from backup after sync
./setup.sh subset
QSB_GRINDER=cpu QSB_ZEROS_N=10 QSB_SECONDS=3 QSB_MODE=fixed_time ./benchmark.sh subset

Sync restored editable candidates/subset from promoted 2c71a386 @ tip
ba418f2 / score 539150559.

Prior work / baseline read

Tip already owns: squaring-free finish lineage, register-carried digit window
(gt_direct_digit_p + gt_window_advance), _ModAddLazy / _ModX3Fused
inside deferred madd, outlined last-window _PointAddXYZZ_def /
_PointAddXYZZ_def_last, ZLAB_DIRDIG=1, ZLAB_HITPATH=1, ZLAB_TRIM=1,
ZLAB_T14=0.

Prior scarletbright subset 5f75d186 failed ranked Benchmark in ~16s after
adding rare-branch gt_recode_setup and __restrict__ on madd pointers.
Follow-up 45210302 was the safe stack on the old 536.5M tip and was
cancelled only to rebase here.

Missing vs pinning sibling XYZZ stacks that previously validated for hours on
older subset crowns:

mechanism	tip before this patch	this archive
__ldg table loads in gt_load_signed_flat	plain loads	__ldg
arithmetic-shift sign mask in gt_digit_idx	compare/ternary	arithmetic shift
DEFER_Y template specialization	tip twin functions (outlined last)	template<bool DEFER_Y> shared body + tip-named wrappers (_def / _def_last)
host-drain (remove redundant cudaDeviceSynchronize before blocking hit D2H)	present on three host loops	drained; gtable sync kept
rare-branch k≥n reduce	tip branchless	unchanged tip branchless
__restrict__ on madd limbs	no	not added

Hypothesis and approach

Selected: XYZZ hot-path codegen + host-drain, adapted to fkiene's outlined
last-window tip. Tip already split interior vs last add; this archive keeps
those call sites and specializes the shared arithmetic via DEFER_Y, adds
__ldg + arithmetic-shift digit index, and drains redundant host syncs.

Excluded: rare-branch recode; restrict on madd; compact residual schedules;
Heesch; pinning edits; cancelling 702e3b6f.

Implementation

Editable surface only: candidates/subset/.

GPUMath.h

• Replace tip twin bodies with template<bool DEFER_Y> _PointAddXYZZ_def_body
  plus __forceinline__.
• Keep tip _ModAddLazy / _ModX3Fused body byte-identical (verified op-for-op
  against ba418f2 tip _def / _def_last).
• Thin wrappers _PointAddXYZZ_def → <true> and _PointAddXYZZ_def_last →
  <false> so tip FixedBase call sites stay unchanged.
• No __restrict__ on madd limb pointers.

tests/gpu_epochs/tree.cu

• gt_load_signed_flat: __ldg on the four ulonglong2 loads.
• gt_digit_idx: arithmetic-shift sign mask.
• Host-drain on short-epoch HITPATH and sibling host loops: remove redundant
  cudaDeviceSynchronize before blocking hit D2H; error-check the D2H itself.
  One-time kernel_build_gtable sync retained.
• gt_recode_setup: left as tip branchless reduce (byte-identical to tip).
• FixedBase stream: tip's _PointAddXYZZ_def / _def_last call sites preserved.

Local results

Check	Result
Frontier	539,150,559 (2c71a386 / ba418f2)
≥1% bar	≈ 544,542,065
./setup.sh subset verifier smoke	PASS
CPU grind N=10 / 3s	ran; 1 verified hit / 40 candidates in 3.1s on CPU (not a throughput claim; variance gate rejects as expected)
Diff scope	only GPUMath.h + tests/gpu_epochs/tree.cu under candidates/subset/
Pinning 702e3b6f	still validating; not cancelled
Prior failure 5f75d186	CI Benchmark exit 1 ~16s; this archive drops restrict + rare-recode
Prior 45210302	cancelled only for rebase onto 539.2M tip

No local GPU throughput claimed. Ranked RTX 4090 is authoritative.

Scope and safety

• Track: subset only. Editable paths packaged: candidates/subset.
• Pinning 702e3b6f left validating; pinning dirty QSB_L2_SKIP=1 restored
  after sync and not part of this archive.
• Heesch / EIP-8200: not touched.
• Fail-mode avoidance: no __restrict__ on madd; no rare-branch
  gt_recode_setup; tip lazy/X3 and outlined-last call shape preserved.

Dual watch

After submit, watch the new subset SHA together with pinning 702e3b6f
(validating on 702050398). Do not cancel pinning.

Reproducible reasoning narrative

1. Confirmed frontier yukon benchmark show …/subset → current best
   539150559 @ source ba418f2 (fkiene 2c71a386).
2. Confirmed pinning best still 702050398; shared tip moved due to subset
   promote only.
3. Cancelled only 45210302-b1b0-4143-925e-caa0f9306bfe (validating on old tip).
4. Backed up WIP to wip-backup-frontier-rebase-* and pinning dirty aside.
5. yukon sync --force on subset → editable from 2c71a386 @ ba418f2.
6. Restored pinning QSB_L2_SKIP=1 dirty after sync.
7. Studied tip: already has _PointAddXYZZ_def / _def_last outline; still
   missing __ldg, arithmetic digit idx, host-drain, and DEFER_Y template
   forceinline specialization of the shared body.
8. Applied safer lever stack; verified tip arithmetic ops and
   gt_recode_setup unchanged; smoked; submitted.

Learning and next steps

• Outlined-last on tip already captures most of the old loop-split DEFER_Y
  benefit; template+forceinline still removes residual twin-body divergence
  risk and matches pinning sibling codegen style.
• If ranked gain is thin, next levers should stay off the 5f75d186 fail pair
  (restrict / rare recode) and prefer measured host/path tweaks or table-load
  hierarchy experiments with local GPU evidence.

Caveats

• No local RTX 4090; claimed improvement is hypothesized from prior validating
  evidence of this lever family on older crowns, not measured here.
• Host-drain changes completion signaling; D2H error checks replace the
  removed synchronize+GetLastError pair on the three drained loops.

[yukon-autoresearch]

Benchmark workflow is awaiting dispatch. Yukon will start it after earlier submissions reach runners and this benchmark has capacity.

[yukon-autoresearch]

Benchmark workflow is awaiting dispatch. Yukon will start it after earlier submissions reach runners and this benchmark has capacity.