allaun/Research-Stack
1
0
Fork 0
mirror of https://github.com/allaunthefox/Research-Stack.git synced 2026-07-31 03:05:21 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions
703 commits 18 branches 0 tags 425 MiB
Commit graph

276 commits

Author SHA1 Message Date
Brandon Schneider
ba1bf871f8 feat(infra): per-tier device limitations for Ray scheduling 
DeviceLimitations dataclass with hard constraints per tier:
  GPU_CUDA:     1GB payload, 16 concurrent, 60s, NVENC, 12GB VRAM
  GPU_VAAPI:    512MB payload, 8 concurrent, 60s, VAAPI HW
  GPU_APU:      256MB payload, 4 concurrent, 30s, shared DDR
  CPU_FFMPEG:   128MB payload, 2 concurrent, 120s, software
  BATCH:        64MB payload, 1 concurrent, 6h, 2000 min/month
  ETHERNET:     1400B payload, 1 concurrent, 1s, virtio-net
  FRAMEBUFFER:  1.5MB payload, 1 concurrent, 100ms, DMA only
  WASM:         1KB payload, 1 concurrent, 10ms, 100K req/day
  DSP:          16KB payload, 1 concurrent, 5s, FFT only
  ESP32:        2KB payload, 1 concurrent, 100ms, Q0_16 scalar

get_limitations(caps) returns actual hardware-aware limits
(vram override, framebuffer capacity, memory override)
 2026-05-30 20:44:31 -05:00
Brandon Schneider
cadb38cc1b feat(infra): add AMD VAAPI + FLAC DSP to FrameDispatcher 
FrameDispatcher now routes 6 tags:
  TAG_STRAND(0x01)  → BraidBackend (VCN compute)
  TAG_CROSSING(0x02) → BraidBackend (VCN compute)
  TAG_PIST(0x03)    → BraidBackend (VCN compute)
  TAG_LUPINE(0x04)  → CUDABackend (NVIDIA CUDA)
  TAG_VAAPI(0x05)   → VAAPIBackend (AMD/Intel VA-API)  ← NEW
  TAG_FLAC(0x06)    → FLACBackend (PipeWire/FLAC DSP)  ← NEW

New backends:
  - VAAPIBackend/LocalVAAPIBackend: AMD/Intel hardware encode/decode
  - FLACBackend/LocalFLACBackend: FFT spectral analysis, centroid, RMS
  - RayVAAPIBackend: Ray actor for VA-API operations
  - SyncVAAPIWrapper/SyncFLACWrapper: sync bridges for FrameDispatcher

Capability probe: DSP tier(1) added between FRAMEBUFFER(2) and ESP32(0)
 2026-05-30 20:35:42 -05:00
Brandon Schneider
5320a08105 feat(infra): integrate edge WASM and GitHub batch compute tiers 
- Update device_capability_probe.py to add BATCH and WASM tiers and fix a NameError bug on has_virtio_net.
- Build Cloudflare Workers WASM compilation and JS fetch handler in 4-Infrastructure/cloudflare/ executing trinary VM steps.
- Create GitHub Actions batch_compute.yml workflow to harvest runner minutes.
- Keep 4-Infrastructure/AGENTS.md updated with the WASM core library anchor.

Build: 3313 jobs, 0 errors (lake build)
 2026-05-30 20:08:31 -05:00
Brandon Schneider
51664eb060 docs(infra): plan — mesh networking layers over Ray 
Architectural plan for Tailscale mesh + Ray + VCN + compute surfaces.

5-layer stack:
  1. Tailscale connects everything (WireGuard, DERP relay)
  2. Ray schedules work across the mesh
  3. VCN compresses data (50x bandwidth reduction)
  4. FrameDispatcher routes by tag
  5. Compute surfaces execute (GPU/CPU/Ethernet/framebuffer/MCU)

Key insight: every device in the Tailscale mesh is a potential
compute node. Framebuffer and Ethernet surfaces turn devices
that "cant run Ray" into compute participants.

4 phases:
  1. Ray over Tailscale (mostly done)
  2. Multi-tier scheduling (probe done, placement pending)
  3. Framebuffer + Ethernet integration (host-side pending)
  4. Edge devices (ESP32, 1-Wire sensors)
 2026-05-30 20:03:57 -05:00
Brandon Schneider
d6fc9cfe6d feat(infra): add ETHERNET compute tier for virtio-net PistPacket DMA 
New tier between CPU_FFMPEG and FRAMEBUFFER:
  GPU_CUDA(7) > GPU_VAAPI(6) > GPU_APU(5) > CPU_FFMPEG(4) >
  ETHERNET(3) > FRAMEBUFFER(2) > ESP32(1) > RELAY(0)

- _detect_virtio_net(): probes /sys/class/net for virtio driver (0x1af4)
- PistPacket computation via TX/RX descriptor rings
- Host vhost-user backend does matrix transforms
- CRC32 hardware offload = witness verification
- Works in any VM with network (even without framebuffer)
 2026-05-30 20:02:06 -05:00
Brandon Schneider
66abf92214 fix(infra): handle Cirrus Logic virtual VGA and DRM naming edge cases 
- Add Cirrus Logic (0x1013) and virtio (0x1af4) to vendor map
- Fix DRM card parsing for names like "card0-VGA-1"
- Virtual GPUs (cirrus, virtio) never classified as discrete
- Virtual GPUs skip VA-API tier, fall to FRAMEBUFFER

Racknerd microVM (2vCPU, 715MB, Cirrus VGA) correctly classified as
FRAMEBUFFER tier: 1024x768 @ 16bpp = 1.57 MB DMA backplane.
 2026-05-30 19:59:38 -05:00
Brandon Schneider
16101a787f feat(infra): device capability probe with framebuffer fallback 
device_capability_probe.py: classify every device into a compute tier.

Tiers (highest to lowest):
  GPU_CUDA    — NVIDIA discrete + CUDA (NVENC, Ray GPU worker)
  GPU_VAAPI   — AMD/Intel discrete + VA-API (hardware encode)
  GPU_APU     — AMD integrated, yuvj420p, bandwidth-optimized
  CPU_FFMPEG  — Software encode only (libx264)
  FRAMEBUFFER — /dev/fb0 DMA backplane (8.29 MB/frame at 1080p)
  ESP32       — MCU, Q0_16 scalar in FreeRTOS idle hook
  RELAY       — Network only, no compute
  OFFLINE     — Unreachable

Features:
  - Multi-GPU DRM render node scanning (card0=AMD, card1=NVIDIA)
  - APU vs dGPU classification via device name + VRAM heuristics
  - Framebuffer detection with /sys/class/graphics/fb0 resolution
  - Ray scheduling helpers (get_ray_placement_strategy)
  - Cluster probe via SSH
  - JSON + human-readable output
 2026-05-30 19:57:07 -05:00
Brandon Schneider
cf907e2835 feat(infra): add QEMU graphics framebuffer packing shim and spec 
- Add qemu_framebuffer_packer.py supporting ARGB8888/RGB24 raw matrix mapping
- Implement zero-copy mmap write/read interface to /dev/fb0 with signature headers
- Document the QEMU graphics framebuffer backplane in Section 11 of spec
- Update 4-Infrastructure/AGENTS.md and walkthrough.md documentation
- Verify syntax and workspace compilation baseline status

Build: 3313 jobs, 0 errors (lake build)
 2026-05-30 19:49:25 -05:00
Brandon Schneider
850f644e0f feat(infra): Ray VCN bridge — FrameDispatcher over Ray transport 
ray_vcn_bridge.py: Ray transport for the VCN-LUPINE bridge.
  Replaces GPUNodeConnection TCP/MKV transport with Ray ObjectRef.
  FrameDispatcher, BraidBackend, CUDABackend are unchanged — only
  the wire between daemon and GPU node changes.

  - RayBraidBackend: compute actor matching VCNBraidBackend pattern
  - RayCUDABackend: GPU actor with /dev/dri (Mesa, no NVIDIA plugin)
  - RayVCNBridge: full bridge as Ray actor (replaces daemon)
  - RayGPUNodeConnection: drop-in for GPUNodeConnection
  - SyncBraidWrapper/SyncCUDAWrapper: bridge Ray actors to sync interface

  STRAND 42B → 63B, CROSSING 42B → 22B, PIST 24B → 57B
  Batch 10: 5ms (0.5ms/frame), 10/10 non-empty
 2026-05-30 19:48:34 -05:00
Brandon Schneider
30d4772a3f feat(infra): support heterogeneous environments in video compute decoder 
- Implement dynamic resolution and format probing using ffprobe inside decode_frames
- Eliminate hardcoded YUV420 frame size slicing during video file readback
- Standardize NVIDIA hardware config to 8-bit full-range yuv444p to keep byte layout unified
- Verify Python compilation and Lean workspace integrity checks

Build: 3313 jobs, 0 errors (lake build)
 2026-05-30 19:48:09 -05:00
Brandon Schneider
2bce7abaa9 feat(infra): optimize AMD APU/iGPU lossless pipeline targeting H.265 cores 
- Add detection for integrated AMD graphics (APUs/iGPUs) based on hardware model name
- Configure UMA-friendly full-range yuvj420p format to reduce system memory bandwidth footprint by 50%
- Force lossless constant QP (-qp 0) and full PC range to prevent clamping loss
- Re-run syntax checks and Lean compiler verification tests

Build: 3313 jobs, 0 errors (lake build)
 2026-05-30 19:46:34 -05:00
Brandon Schneider
6ea34fbae6 feat(infra): add GPU-specific math optimization loader for H.265 VCN/NVENC 
- Implement MathOptimizationLoad dataclass to represent GPU packing configurations
- Update probe_vcn_capabilities to resolve optimizations for NVIDIA/AMD/Intel GPUs
- Extend compute_frame_size to support yuvj420p, 10-bit YUV, and YUV444p
- Propagate optimized pixel formats into select_optimal_resolution and spec
- Update _build_ffmpeg_cmd to inject lossless/zero-latency options and HEVC/H.265 metadata SEI NAL parameters
- Update 4-Infrastructure/AGENTS.md documentation

Build: 3313 jobs, 0 errors (lake build)
 2026-05-30 19:45:09 -05:00
Brandon Schneider
63093f56da feat(infra): Ray VCN transport + cluster restoration 
- ray_vcn_transport.py: @ray.remote wrappers for braid VCN encode/decode
  - Distributed encode on CPU workers, compute on GPU workers
  - RayVCNTransport actor with frame counter + ObjectRef storage
  - FAMM-gated encode task, batch encode/decode helpers
  - 20 strands in 576ms (28.8ms/strand), 20/20 CRC ok

- raycluster.yaml: KubeRay cluster on qfox-1
  - Head + CPU worker + GPU worker (RTX 4070 SUPER via /dev/dri)
  - No NVIDIA device plugin — Mesa direct device access
  - Tolerations for desktop taint on qfox-1
  - num-gpus instead of custom GPU resource

- fix-nftables-k3s.sh: nftables forward rules for flannel/cni0
  - nftables default policy=drop blocks pod-to-pod networking
  - systemd service nftables-k3s-fix for persistence

- KubeRay operator moved to nixos (control plane can reach API server)
- FFmpeg 8.0 + reedsolo installed in Ray head pod via conda
 2026-05-30 19:42:39 -05:00
Brandon Schneider
61b38ca697 refactor(infra): optimize YUV420 frame packing using numpy 
- Vectorized create_yuv420_frame when numpy is available to eliminate the 500k-iteration scalar Python loop.
- Pre-filled memoryview slice buffers in the fallback path.
- Updated 4-Infrastructure/AGENTS.md to document the optimization.

Build: 3313 jobs, 0 errors (lake build)
 2026-05-30 19:32:11 -05:00
Brandon Schneider
30d8c56158 feat(infra): improve RRC Ray Layer Tagger and align registry 
Improved rrc_ray_tagger.py with prioritized source name-based variant matching, corrected NetworkRayReceipt (3 variants, 67us) and BurgersRGSolver (5 variants) shapes, fixed Hopf-Cole fallback bug using string normalization, dynamically deduced workspace root path, and quarantined phase_update due to adversarial review falsification. Registered anchor in 4-Infrastructure/AGENTS.md.

Build: 3313 jobs, 0 errors (lake build Compiler)
 2026-05-30 19:18:02 -05:00
Brandon Schneider
4965029758 feat: integrate May 2026 math papers into Research Stack 
1. Singer Sidon Sets (2605.03274):
   - New SidonSets.lean: IsSidon, IsSidonMod, IsIntervalSidon, h(N)
   - 5 fully proved lemmas, 13 sorry with TODO(lean-port)
   - GoldenRatioSeparation.lean: singer_density_lt_golden (proved)
   - lake build: 3303 jobs, 0 errors

2. Hexagonal lattice + RG (2605.09974):
   - New test_hexagonal_lattice_rg() in unified_rg_tests.py
   - Avila's global theory exact phase diagram
   - RG confirms localized/extended regimes
   - Fractal dimension: extended→1, critical→0.5, localized→0
   - 7 tests, all pass

3. Burgers + Hopf-Cole + Fokas (2605.11788):
   - Added solve_heat_fokas() — unified transform method
   - Added solve_burgers_fokas() — full Burgers via Hopf-Cole + Fokas
   - Added solve_heat_fourier_series() — comparison solver
   - Fokas converges in ~64 quadrature points vs Fourier 2000 terms
   - Hopf-Cole FFT: 8-208x faster than finite differences
 2026-05-30 18:16:57 -05:00
Brandon Schneider
a53e023cbe feat: Hopf-Cole exact solver for 1D Burgers — 151x speedup 
Hopf-Cole transformation maps Burgers to heat equation:
  u = -2v * d(ln ψ)/dx
  dψ/dt = v * d²ψ/dx² (exact via FFT)

Benchmark results:
  N=512,  v=0.01: 1.45x speedup
  N=1024, v=0.01: 83.4x speedup
  N=2048, v=0.01: 151.3x speedup

Key insight from adversarial review:
  - RG assumption (nonlinear term vanishes) is FALSE
  - But 1D Burgers IS integrable via Hopf-Cole
  - Exact solution in O(N log N), no time stepping
  - The 'insultingly easy' regime exists — just not via RG

This is the exact solution the agents found when they
broke the RG fixed point assumption.
 2026-05-30 17:36:12 -05:00
Brandon Schneider
c556d64ae0 feat: QEMU compute surfaces — virtio-crypto + ivshmem 
virtio_crypto_transform.py:
- VirtioCryptoSession: HASH session (SHA-256, SHA-512, MD5)
- VirtioCryptoHashTransform: encode as HASH request, produce receipt
- Receipt: {schema, transform_type, algo, payload_bytes, result_hex, witness_hash}
- Wire-format structs: CtrlHdr(20B), HashSessionPara(8B), HashDataReq(28B)
- RFC 6234 test vectors: all pass

ivshmem_client.py:
- IvshmemClient: mmap /dev/shm/ivshmem_bar0
- IvshmemRing: doorbell notification
- IvshmemTransform: write payload, ring doorbell, produce receipt
- Receipt: {schema, transform_type: shared_memory, offset, length, witness_hash}
- Memory layout: registers 0x0000, metadata 0x10000, data 0x20000
- /dev/shm fallback test: verified
 2026-05-30 17:35:24 -05:00
Brandon Schneider
f63e4b5179 feat(infra): virtio-net ring as compute pipeline 
Add virtio_net_transform.py: three Class-1 computation primitives via
virtio-net TX/RX rings — zero backend code changes needed.

  1. HASH_REPORT — host writes Toeplitz RSS hash into RX header
     (virtio_net_hdr_v1_hash.hash_value return channel)
  2. TSO gso_size — host splits large buffer via TCP segmentation offload
     (spatial partition into N × gso_size chunks)
  3. MRG_RXBUF — host merges multiple RX buffers (aggregation primitive)

Structs: VirtioNetHdr (12B), VirtioNetHdrHash (20B), VringDesc (16B).
Receipt schema: virtio_transform_receipt_v1 with CRC32 witness_hash.

The copy-if filter (skip zero deltas, process non-zeros) maps directly
onto the HASH_REPORT return channel: delta=0 → hash skip, delta≠0 →
hash_as_function_of_payload. This is the ambient compute model:
any QEMU/firecracker microVM is already a computation device without
knowing it.

Build: 3313 jobs, 0 errors (lake build)
Tools: glslang, spirv-as, spirv-dis (native); tint from nixpkgs for WGSL
 2026-05-30 16:54:09 -05:00
Brandon Schneider
9c5fe97dc1 feat(infra): SPIR-V packet generator and WGSL scar filter shader 
Add spirv_packet_generator.py: reads SPIR-V assembly, applies copy-if
optimization (OpPhi→OpSelect transform), and emits JSON packet descriptors
with the 5 OpPhi-derived fields (type_id, cond_id, true_val_id,
false_val_id, result_id) that fully specify the packet layout.

Add burgers_scar_filter.wgsl: 291-line WGSL compute shader for spectral
scar filtering in 2D Burgers RG solver. Uses three copy-if patterns:
  1. scar_pressure > threshold → apply hyperviscosity damping
  2. |kx| > k_cut || |ky| > k_cut → zero (dealiasing)
  3. factor < 0.999 → multiply velocity components

Also fix spirv_copy_if_optimizer.py: OpSelect now uses phi_instr.args[0]
(type operand) as its type, instead of compute_instr.result_id. This
produces structurally correct SPIR-V where the result type matches the
OpSelect opcode layout.

Build: 3313 jobs, 0 errors (lake build)
Tools: glslang, spirv-as, spirv-dis (native); tint from nixpkgs for WGSL
 2026-05-30 16:40:58 -05:00
Brandon Schneider
cfd43e1e95 feat(codec): extend BraidDiatCodec with BraidDiatFrame encoder/decoder 
- BraidDiatCodec.lean: BraidDiatFrame now handles encode/decode of full
  SpherionState × BraidReceipt with 256-bit header and variable mountain list
- braid_diat_codec.py: Python extraction updated to match, benchmark artifact
  at shared-data/artifacts/braid_diat_codec_benchmark.json (714B avg vs
  messagepack 1748B avg)

Build: lake build Compiler 3313 jobs, 0 errors
 2026-05-30 16:23:41 -05:00
Brandon Schneider
4df6997b51 docs(kube): update infra docs, RayCluster manifest with nightly GPU images 
- infrastructure-status.md: rewrite with current cluster topology (cupfox control-plane,
  neon-64gb/racknerd/steamdeck workers), RayCluster status, Garage storage,
  Caddy edge, open issues
- k3s-cluster-setup.md: fix steamdeck hardware specs (8 vCPU, 14.5 GB RAM)
- raycluster.yaml: upgrade to rayproject/ray:nightly-py313-gpu (multi-arch amd64+arm64),
  add gpu-workers group targeting neon-64gb, add arm64-workers for neon-64gb CPU
- README.md: update build job count (3460 → 3313, verified)

Build: lake build Compiler 3313 jobs, 0 errors
 2026-05-30 16:23:13 -05:00
Brandon Schneider
81d4338627 feat: ARM64 copy-if optimizer — branches to CSEL 
Transforms branch patterns to ARM64 conditional selects:
  Before: CMP + BEQ + compute + B + MOV = 5-47 cycles
  After: CMP + compute + CSEL = 4-6 cycles

ARM64 CSEL instruction:
  CSEL Xd, Xn, Xm, cond
  - Single cycle on most ARM64 processors
  - No branch prediction penalty
  - No pipeline flush on mispredict

Pattern detection:
  - CMP + BEQ/BNE/B.LT/etc
  - True block: 1-3 compute instructions + B
  - False block: single MOV
  - Merge point

Same pattern as:
  - SPIR-V OpSelect (GPU shaders)
  - VCN delta+RLE (3.3x)
  - QR spatial hash (2.18x)
  - Lean CopyIfTactic (2.7x)

Works on ARM64 assembly from GCC/LLVM/Rust.
No compiler fork needed — post-processing pass.

Targets: Neon-64GB (18 vCPU ARM64 EPYC)
 2026-05-30 15:59:24 -05:00
Brandon Schneider
2e15c7c0a5 feat: SPIR-V copy-if optimizer — skip zero deltas in GPU shaders 
Transforms branch-based patterns to OpSelect:
  Before: 3 blocks, OpBranchConditional, OpPhi
  After: 1 block, OpSelect (single-cycle on most GPUs)

Pattern detection:
  - OpSelectionMerge + OpBranchConditional
  - True block: single compute + OpBranch
  - False block: empty (just OpBranch)
  - Merge block: OpPhi merging true/false values

Transformation:
  - Remove SelectionMerge + BranchConditional
  - Inline compute instruction
  - Replace OpPhi with OpSelect
  - Collapse 3 blocks to 1

Same pattern as:
  - VCN delta+RLE (3.3x): skip zero bytes
  - QR spatial hash (2.18x): skip non-neighbors
  - Lean compilation (2.7x): skip trivial theorems
  - Spatial hash (86.5% cache hit): skip empty cells

Driver-agnostic: works at SPIR-V level before Mesa.
No Mesa fork required. No NIR pass needed.
 2026-05-30 15:57:08 -05:00
Brandon Schneider
d5428a8950 feat: QR spatial hash integration — 2.18x speedup 
Cache-friendly Householder QR via Morton-code spatial hash:
- When adding column, only apply reflections to 3x3x3 neighborhood
- Reduces per-update from O(n) to O(27) per column
- 50x50 matrix, 500 updates: 2.18x faster than naive

Naive: 0.124ms/update
Spatial: 0.057ms/update
Speedup: 2.18x

Key insight: Morton code ordering means nearby columns in 3D
are nearby in memory → cache-friendly access → fewer misses.

This completes all 4 next steps:
1. ✅ O_AMMR_QRNode wired into BraidDiatFrame (already done)
2. ✅ O_AMMR_valid strengthened with residual bounds (NS_MD.lean)
3. ✅ Hash benchmark: Morton wins (86.5% cache hit rate)
4. ✅ QR spatial hash: 2.18x speedup
 2026-05-30 15:30:06 -05:00
Brandon Schneider
bdc227459a feat: O_AMMR_valid strengthened + hash benchmark complete 
NS_MD.lean:
- Added QRResidualWitness structure (Q16_16 fixed-point)
- Added residual_bound_ok, basis_size_ok, orthogonality_ok predicates
- Extended O_AMMR_Node with qr_witness field
- Strengthened O_AMMR_valid: 4 conjuncts (admission + residual + basis + ortho)
- lake build: 3300 jobs, 0 errors

hash_benchmark.py (240 data points):
- Hilbert vs Morton vs xxHash
- 5 grid sizes (16^3 to 256^3), 4 trace sizes, 4 patterns

Key findings:
  Morton: 86.5% cache hit rate, 1.08µs p50, 0.512 locality
  xxHash: 30.3% cache hit rate, 0.96µs p50, 0.342 locality
  Hilbert: 27.6% cache hit rate, 2.29µs p50, 0.833 locality

Morton wins overall for spatial hash grids.
 2026-05-30 15:15:33 -05:00
Brandon Schneider
c6011dbbdf feat: wire BraidDiatCodec into FAMM transport 
BraidDiatCodec (714 bytes avg) imported alongside VCN encoder.
When available, can replace Delta+RLE for braid data encoding.

Benchmark (from braid_diat_codec_benchmark.json):
  BraidDiat: 0.029ms encode, 0.034ms decode, 714 bytes
  MessagePack: 0.103ms encode, 0.002ms decode, 1748 bytes
  Cap'n Proto: 0.002ms encode, 0.0001ms decode, 29 bytes

BraidDiatCodec is 2.5x smaller than MessagePack and encodes
braid data natively (Q0_2 fields, Mountain packed, MMR state).
 2026-05-30 14:36:16 -05:00
Brandon Schneider
c64e4f15eb docs: update AGENTS.md — BraidDiatCodec shim + bridge modules 
4-Infrastructure: add braid_diat_codec.py to stack-solidification anchors
Semantics: document BraidSpherionBridge admits (IntNodeToPhaseVec linearity, receipt_correspondence)
 2026-05-30 13:30:55 -05:00
Brandon Schneider
e797f06bd0 feat(shim): BraidDiatCodec Python extraction 
Python implementation of the 4-layer BraidDiatCodec:
- ChiralityDIAT encode/decode (64-bit slot)
- MountainPacked from_mountain/to_mountain
- BraidResidualPacked from_bracket/to_bracket (Q0_2 packing)
- BraidDiatFrame encode/decode

Benchmark: braid_diat (714B avg) vs messagepack (1748B avg)
on synthetic MMR/spike train frames.
 2026-05-30 13:30:01 -05:00
Brandon Schneider
ba203ca971 feat: spatial hash grid — GPU-style particle physics (Python + FPGA) 
Ported from ScaleSpaceSynth (WebGPU particle simulator):
- 64×64×64 spatial hash, 32 particles/cell, lock-free insertion
- Curl noise: divergence-free 3D turbulence
- Pairwise forces: attractive (ratio>0.15) + repulsive (ratio<=0.15)
- Trilinear density interpolation
- HalfLife particle lifecycle
- Q16_16 encode/decode for VCN transport

Python (spatial_hash_grid.py): 6/6 tests pass
  10K particles, neighbor query, forces, 100 sim steps, curl noise verified

FPGA (spatial_hash_bram.v):
  16×16×16 grid, dual-port BRAM, 27-cycle neighbor scan
  Density → voltage mode selector (STORE/COMPUTE/APPROX/MORPHIC)
  Integrated into research_stack_top.v

Same pattern as ScaleSpaceSynth GPU:
  GPU: atomicAdd for lock-free cell assignment
  FPGA: BRAM read-modify-write for cell assignment
  Ray: content-addressed ObjectRef for lock-free reads
  All: partition space → compute density → find structure at multiple scales
 2026-05-30 01:25:19 -05:00
Brandon Schneider
df6ea7ba15 feat: GCCL + WaveProbe + MetaProbe + delta compression 
Ports Lean formalization to Python:
- GCCL: LawAxis, PromotionRung, Decision, ScaleBand, Receipt, Wrapper, Transition
- gcclSwapGate: accept if new cost < old cost (from MassNumber.lean)
- fammRouteGate: route mass <= stress mass within thermal budget
- braidTransferGate: delta admissible <= delta risk
- WaveProbe: golden angle sampling (40503 = 1/φ × 65536)
- MetaProbe: probe-but-don't-commit, EXPORT_GRANT
- Delta compression with GCCL gates
- gccl_encode: full GCCL-gated encode pipeline

Tests:
  WaveProbe overlap (identical): 1.0000
  WaveProbe overlap (shifted): 0.6694
  MetaProbe (low residual): EXPORT_GRANT
  MetaProbe (high residual): HOLD
  GCCL transition admissible: True

All Q16_16 arithmetic (no Float in compute paths).
 2026-05-30 01:09:14 -05:00
Brandon Schneider
ebabaa3b6b feat: FAMM-integrated VCN transport (gate-checked encode/decode) 
Ports Lean formalization to Python:
- gateCondition: ||coker(M) residual|| < ε (Q16_16)
- Scar/ScarBundle: pressure + mode per strand
- fammGate: admissibility check on 8-strand state
- eigensolid_converged: verify convergence before transmission
- voltage_mode_from_fd: FD → STORE/COMPUTE/APPROX/MORPHIC
- latency_class: RTT → local/near/far/derp/offline

Pipeline:
  Braid data → FAMM gate → eigensolid check → FD → voltage mode
  → RouteCost latency → VCN encode → SEI receipt with FAMM metadata

Gate behavior:
  - FAMM admissible + eigensolid converged → encode
  - Either fails → reject with scar info, don't encode
  - Receipt includes: claim_boundary, promotion=not_promoted

All Q16_16 arithmetic (no Float in compute paths).
68/68 tests still pass.
 2026-05-30 00:56:43 -05:00
Brandon Schneider
65d0ddb12a fix(deps): update tar to v0.4.46 and refresh npm packages 
Resolves GHSA-9rg2-7cm8 (PAX header desynchronization) in:
- 2-Search-Space/search/stract/Cargo.lock
- 4-Infrastructure/servo-fetch/Cargo.lock

Note: @ai-sdk/provider-utils in dify-ai-provider nested dep (3.0.25)
has no fix available yet; tracked in GHSA-866g-f22w-33x8
 2026-05-30 00:37:43 -05:00
Brandon Schneider
e5fb0a5f4d chore: commit accumulated working tree changes 
Lean: update Semantics modules, add new numerics/physics data files
Hardware: update FPGA bitstreams (tangnano9k_uart_loopback)
Infra: k3s-flake tests, netcup-vps configuration, VCN compute substrate
Docs: ARCHITECTURE, specs, citation updates
 2026-05-30 00:10:02 -05:00
Brandon Schneider
69e5b31c4f docs(infra): add Ollama inference serving to k3s-cluster-setup 
Add inference serving section documenting:
- Ollama host-level deployment on Neon-64GB (bypasses KServe RAM limits)
- Caddy reverse proxy on racknerd (:8443 → 100.64.19.78:11434)
- Troubleshooting: autosave.json stale config, passt port interception
- curl/wget verification commands

Fixes 502 Bad Gateway from Caddy (--resume loading stale autosave.json).
 2026-05-30 00:02:49 -05:00
Brandon Schneider
9f304abab0 feat: fractal dimension — DBC algorithm (Python + FPGA) 
Paper: 'Ultra-fast computation of fractal dimension for RGB images'
  (Pattern Analysis and Applications, 2025)

Python (fractal_dimension.py):
- DBC algorithm with numpy vectorization (29x faster than scalar)
- fd_compress_hint: FD → voltage mode (STORE/COMPUTE/APPROX/MORPHIC)
- 7/7 tests pass (Sierpinski, random, gradient, checkerboard, fBm, constant, RGB)
- Q16_16 integer arithmetic internally

FPGA (fractal_box_counter.v + fractal_fd_selector.v):
- 5-state FSM: IDLE → COLLECT → FINALIZE → STORE_LOG → REGRESS → DONE
- 8 power-of-two scales (2, 4, 8, ..., 256)
- Linear regression via Q16_16 64-bit arithmetic
- FD clamped to [1.0, 3.0] in Q16_16
- Selector: FD < 2.3 → STORE, < 2.6 → COMPUTE, < 2.9 → APPROX, >= 2.9 → MORPHIC
- Integrated into research_stack_top.v

FD drives adaptive compression:
  Low FD (smooth) → STORE mode (minimal compression)
  High FD (rough) → MORPHIC mode (aggressive compression)
 2026-05-29 20:45:21 -05:00
Brandon Schneider
c43f590fc0 feat(infra): add framebuffer/DRM/VirtIO GPU 3D acceleration 
hardware.virtio.enable + guestAgent (QEMU guest agent for SCP).
services.xserver with virtiogpu + fbdev + vmware video drivers.
services.spice.enable with vdagent.

Packages: mesa, libGL, libGLU, virglrenderer (VirtIO-gpu 3D),
swiftshader (software Vulkan), libva (VA-API), spice, spice-gtk,
xorg.libX11/ext/render, xf86-video-vMware/fbdev.

Kernel modules: virtio-gpu, drm, drm_kms_helper, drm_shmem_helper, ttm.
 2026-05-29 15:09:58 -05:00
Brandon Schneider
b93069406d feat(infra): mirror Debian kernel modules + SCP/KVM guest support 
Kernel modules: virtio_blk/scsi/net/balloon/serial/gpu/dma_buf,
drm/drm_kms_helper/drm_shmem_helper, xhci_pci/usbcore/usbhid,
scsi_mod/sr_mod/cdrom, ext4/mbcache/jbd2, efi_pstore/efivarfs,
qemu_fw_cfg, ip_tables/x_tables.

QEMU guest agent + SPICE vdagent for netcup SCP control panel
(graceful shutdown, VNC console, SPICE display, guest info).

VirtIO drivers for KVM guest environment.

Kernel params: net.ifnames=0, console=tty0, quiet.

Clean install — filesystems will be partitioned during NixOS install.
 2026-05-29 15:07:52 -05:00
Brandon Schneider
c2e51aa40a feat(infra): ARM64 performance tuning for netcup-vps 
Per-Ampere-tunable sysctls:
- vm.swappiness=10, dirty_ratio=60, dirty_background_ratio=15
- Transparent hugepages (1024 huge + 1024 overcommit)
- ARM64 shmmax/shmall/shmni raised for Julia/PETSc (32GB shared mem)
- zone_reclaim_mode=0 (NUMA-aware, no local-only allocation)
- fs.file-max/nr_open=524288, inotify.max_user_watches=524288
- Network buffers tuned for LSP connections (16MB rmem/wmem, TCP FastOpen)
- POSIX msg queues raised

PAM loginLimits: nofile/memlock unlimited.
Environment: OPENBLAS_NUM_THREADS=16, PETSC_OPTIONS=-matpthread,
JULIA_NUM_THREADS=16, OMP_NUM_THREADS=16.

New packages: numactl for NUMA affinity control.
 2026-05-29 14:59:40 -05:00
Brandon Schneider
92a5a6f332 feat(infra): enable btrfs kernel support and tools on netcup-vps 
btrfs-progs, btrfs-heatmap, btrfs-static.
boot.kernelModules += btrfs, boot.supportedFilesystems += btrfs.
grub.fsTracker enabled for subvolume tracking.
 2026-05-29 14:58:20 -05:00
Brandon Schneider
b7c16e0237 feat(infra): add Jellyfin media server to netcup-vps 
jellyfin, jellyfin-ffmpeg, jellyfin-web packages.
Systemd service on port 8096, discovery on UDP 1900.
Caddy reverse proxy ready for jellyfin when domain is configured.
 2026-05-29 14:57:34 -05:00
Brandon Schneider
09599ca856 feat(infra): add ffmpeg, audio/video DSP packages to netcup-vps 
ffmpeg-full, flac, opus, libvpx, libaom, dav1d, mediainfo, sox,
portaudio, libsndfile, alsa-lib — full audio/DSP pipeline.
 2026-05-29 14:54:09 -05:00
Brandon Schneider
4bc9de1e09 feat(infra): netcup-vps — podman+k3s, PostgreSQL, Caddy, Prometheus, full ARM64 math stack 
Swap Docker for Podman + k3s (single-node server).

New services:
- PostgreSQL 16 with JIT + tuned memory (16GB shared_buffers, 48GB cache)
- Caddy reverse proxy (HTTPS for LSP endpoints — needs domain)
- Prometheus node exporter (port 9100)
- NixOS weekly upgrade timer
- Health watchdog (restarts LSP/Ollama if unhealthy)

New packages (ARM64-optimized):
- openblas, blis, lapack — multi-threaded linear algebra
- petsc, slepc — sparse/eigenvalue solvers
- flintqs, pari, gap, singular — number theory / algebra
- symengine — fast C++ symbolic (SymPy backend)
- fftw, suitesparse — FFT and sparse direct solvers
- z3, julia_11 — SMT and JIT numerics

k3s ports: 6443, 2379, 2380
Firewall updated accordingly.
 2026-05-29 14:52:20 -05:00
Brandon Schneider
24d9206857 feat(infra): add Tailscale mesh, tmpfs RAM disks, Nix caching, ENE restore 
- 32GB /tmp and /run/shm tmpfs for fast build scratch
- NixOS cache + Lean community cache substituters
- Tailscale mesh networking (server mode)
- ENE database restore systemd service
- Aggressive parallelism (LAKE_JOBS=16, NIX_BUILD_CORES=16)
- vm.swappiness=10, 1024 hugepages

Packages build passes. System config needs real disk layout from CCP.
 2026-05-29 14:32:56 -05:00
Brandon Schneider
066a96f03f docs(infra): update node topology with rs-vps (netcup ARM64) 2026-05-29 14:26:46 -05:00
Brandon Schneider
750ffea1dd feat(infra): add netcup VPS flake (ARM64, NixOS 25.11) 
Provision: lean-lsp-mcp (8765/8766), pylsp (8767), ollama (11434)
Services: systemd units for Lean 4.19.0, 4.30.0-rc2, Python LSP, Ollama
Packages: elan, uv, texliveFull, octave, typst, docker, nix-ld

Hardware: Ampere ARM64, 64GB RAM, 18 cores, 2TB disk
Build: nix build .#packages.aarch64-linux.lean-lsp-mcp
 2026-05-29 14:26:07 -05:00
Brandon Schneider
af2fa96c35 feat(shim): add pylsp_trivial_detector plugin 
Pre-filter plugin for python-lsp-server that handles trivial changes
instantly (whitespace, comments, docstrings, imports, fast syntax errors)
to reduce full analysis overhead. Mirrors the copy_if pattern from
Semantics.CopyIfTactic for Lean.

Setup: uv tool run --from python-lsp-server[all] --with pylsp-trivial
Laptop: /home/allaun/.local/lib/pylsp-trivial
 2026-05-29 13:51:21 -05:00
Brandon Schneider
7047a770d1 feat: AlphaProof batch mode with copy-if pre-filter 
New CLI mode: python alphaproof_loop.py --batch <lean_dir>

Pipeline:
1. Pre-filter scans Lean codebase (11,434 theorems)
2. Classifies trivial (63.5%) vs non-trivial (36.5%)
3. Prioritizes non-trivial by tactic count + sorry weight
4. Feeds hardest problems to Ollama LLM first
5. Skips trivial theorems entirely (zero deltas)

Usage:
  python alphaproof_loop.py --batch ../../0-Core-Formalism/lean/Semantics/Semantics
  python alphaproof_loop.py --batch . --max-iter 20 --model deepseek-coder-v2:16b

Same pattern as vectorized copy_if:
  - Skip zero deltas → 40x faster (blog post)
  - Skip trivial theorems → 2.7x faster (pre-filter)
  - Focus solver on residuals → fewer iterations
 2026-05-29 02:42:04 -05:00
Brandon Schneider
feebe41d14 feat: Lean proof pre-filter (copy-if pattern for compilation) 
63.5% of theorems are trivial (zero deltas) — solver should skip them.
Estimated speedup: 2.7x (11,434 theorems → 4,170 need solving).

Classifies theorems as trivial/non-trivial:
  Trivial: rfl, decide, trivial, alias, constructor, documented sorry, #eval
  Non-trivial: simp, omega, native_decide, ring, linarith, sorry

Top 10 heaviest modules identified (EntropyMeasures has 152 non-trivial).

Usage:
  python3 lean_proof_prefilter.py <file.lean>
  python3 lean_proof_prefilter.py --scan <dir/>

Same pattern as vectorized copy_if:
  - Blog: skip zero deltas in VPCOMPRESSD → 40x faster
  - Lean: skip trivial theorems in simp → 2.7x faster
  - VCN: skip zero deltas in delta+RLE → 3.3x faster
 2026-05-29 02:34:38 -05:00
Brandon Schneider
7237b2e09b feat: vectorized delta+RLE via copy-if pattern (3.3x faster, 2.5x smaller) 
Inspired by loonatick-src vectorized copy_if analysis:
- VPCOMPRESSD memory-dest = 144 microcode uops (40x bottleneck)
- Register-dest compress + regular store = 10-40x faster

Applied to VCN pipeline:
- numpy vectorized delta (np.diff) + copy-if (nonzero mask)
- RLE on filtered stream = concentrated runs = better compression
- Falls back to scalar if numpy unavailable or data < 1024 bytes

Benchmark (800KB random data):
  Scalar:     132.8ms, 499KB output (0.62 ratio)
  Vectorized:  40.2ms, 200KB output (0.25 ratio)
  Speedup: 3.3x, compression: 2.5x smaller

Wire: delta_rle_encode_vectorized() replaces delta_rle_encode() in pipeline
 2026-05-29 02:24:55 -05:00