feat(autoresearch): add network protocol task, standardize branch naming

New autoresearch task:
- network/protocol: message serialization, tensor wire format, zstd
  compression, CRC32 checksums, auth handshake, discovery packets,
  topology layer lookup and auto-assignment

Branch naming (all 11 tasks):
- Replaced date-tag scheme (autoresearch/<task>/<tag>) with fixed
  branch names matching directory path (autoresearch/backends/cuda,
  autoresearch/models/tts, autoresearch/network/protocol, etc.)
- Updated all program.md files and README.md

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: evilsocket

autoresearch/README.md

@@ -1,6 +1,6 @@
 # Cake Autoresearch
 
-Autonomous optimization of inference throughput across models, backends, and kernels.
+Autonomous optimization of inference throughput across models, backends, kernels, and network.
 
 ## Concept
 
@@ -17,15 +17,18 @@ autoresearch/
 ├── models/                     # Model-specific optimization
 │   ├── text-inference/         # Attention, MLP, cache, transformer blocks
 │   ├── moe/                    # Mixture-of-Experts routing and dispatch
+│   ├── tts/                    # TTS: diffusion head, DDPM, VAE, vocoder, mel
 │   └── image-generation/       # FLUX/Stable Diffusion components
 ├── backends/                   # Backend-specific GPU/CPU optimization
 │   ├── cuda/                   # NVIDIA CUDA kernels (ops.cu, ops.rs)
 │   ├── metal/                  # Apple Metal shaders (ops.msl)
 │   └── vulkan/                 # Vulkan/wgpu compute shaders (ops.wgsl)
-└── kernels/                    # Kernel-specific fused operation optimization
-    ├── attention/              # Scaled dot-product attention
-    ├── fused-ops/              # silu_mul, rms_norm, add3, add_scaled, etc.
-    └── quantization/           # FP8 dequantization and weight preprocessing
+├── kernels/                    # Kernel-specific fused operation optimization
+│   ├── attention/              # Scaled dot-product attention
+│   ├── fused-ops/              # silu_mul, rms_norm, add3, add_scaled, etc.
+│   └── quantization/           # FP8 dequantization and weight preprocessing
+└── network/                    # Distributed inference protocol
+    └── protocol/               # Serialization, topology, auth, discovery
 ```
 
 Each task directory contains:
@@ -43,11 +46,24 @@ Generated at runtime (gitignored):
 | `baseline.txt` | Baseline benchmark score from `prepare.sh` |
 | `experiments.tsv` | Tab-separated experiment log |
 
+## Branching
+
+Each task uses a **fixed branch name** matching its directory path:
+
+```bash
+git checkout -b autoresearch/backends/cuda       # CUDA kernels
+git checkout -b autoresearch/models/tts          # TTS models
+git checkout -b autoresearch/network/protocol    # Network protocol
+# etc.
+```
+
+No date tags — one branch per task. Merge to main when done.
+
 ## Quick Start
 
 ```bash
 # 1. Pick a task
-cd autoresearch/kernels/fused-ops
+cd autoresearch/network/protocol
 
 # 2. Run setup (builds, benchmarks baseline)
 bash prepare.sh
@@ -81,31 +97,39 @@ bash prepare.sh
 
 ### Models
 
-| Task | Target | Primary Metric | Benchmarks |
-|------|--------|----------------|------------|
-| text-inference | Common text model components | Forward pass time | attention, mlp, blocks, cache |
-| moe | MoE routing and expert dispatch | MoE forward time | moe, expert_provider |
-| image-generation | FLUX/SD image pipeline | Component time | flux |
+| Task | Branch | Target | Benchmarks |
+|------|--------|--------|------------|
+| text-inference | `autoresearch/models/text-inference` | Attention, MLP, cache, blocks | attention, mlp, blocks, cache |
+| moe | `autoresearch/models/moe` | MoE routing and dispatch | moe, expert_provider |
+| tts | `autoresearch/models/tts` | Diffusion head, DDPM, VAE, vocoder | prediction_head, ddpm, connectors, mel, wav |
+| image-generation | `autoresearch/models/image-generation` | FLUX/SD pipeline | flux |
 
 ### Backends
 
-| Task | Target | Hardware | Benchmarks |
+| Task | Branch | Hardware | Benchmarks |
 |------|--------|----------|------------|
-| cuda | CUDA kernels (ops.cu) | NVIDIA GPU | fused ops (GPU path) |
-| metal | Metal shaders (ops.msl) | Apple Silicon | fused ops (GPU path) |
-| vulkan | WGSL compute shaders (ops.wgsl) | Vulkan 1.3+ GPU | vulkan benchmarks |
+| cuda | `autoresearch/backends/cuda` | NVIDIA GPU | fused ops (GPU path) |
+| metal | `autoresearch/backends/metal` | Apple Silicon | fused ops (GPU path) |
+| vulkan | `autoresearch/backends/vulkan` | Vulkan 1.3+ GPU | vulkan benchmarks |
 
 ### Kernels
 
-| Task | Target | Primary Metric | Benchmarks |
-|------|--------|----------------|------------|
-| attention | Attention implementations | Attention forward time | attention, linear_attn |
-| fused-ops | Fused activation/norm ops | Per-op latency | bench_utils fused ops |
-| quantization | FP8/GPTQ dequantization | Dequant throughput | quantization, flux dequant |
+| Task | Branch | Target | Benchmarks |
+|------|--------|--------|------------|
+| attention | `autoresearch/kernels/attention` | Attention implementations | attention, linear_attn |
+| fused-ops | `autoresearch/kernels/fused-ops` | Fused activation/norm ops | bench_utils fused ops |
+| quantization | `autoresearch/kernels/quantization` | FP8/GPTQ dequantization | quantization, flux dequant |
+
+### Network
+
+| Task | Branch | Target | Benchmarks |
+|------|--------|--------|------------|
+| protocol | `autoresearch/network/protocol` | Serialization, topology, auth, discovery | protocol, serialization, discovery, topology, auth |
 
 ## Prerequisites
 
 - Rust toolchain (stable)
 - `cargo bench` support (divan benchmarks)
 - For backend tasks: appropriate GPU + feature flags (cuda, metal, vulkan)
+- For TTS tasks: `--features vibevoice,luxtts`
 - No model downloads required — all benchmarks use synthetic tensors

autoresearch/backends/cuda/program.md

@@ -11,9 +11,8 @@ NVIDIA hardware. Lower is better.
 
 ## Setup
 
-1. **Agree on a run tag**: e.g. `mar23`. Branch: `autoresearch/cuda/<tag>`.
-2. **Create the branch**: `git checkout -b autoresearch/cuda/<tag>`
-3. **Read the in-scope files**:
+1. **Create the branch**: `git checkout -b autoresearch/backends/cuda`
+2. **Read the in-scope files**:
    - `cake-core/src/backends/cuda/ops.cu` — Raw CUDA PTX kernels. **Primary target.**
      Contains all fused operation kernels: silu_mul, stable_softplus, rms_norm_gated,
      add_rms_norm, rms_norm_channel, add3, add_scaled, depthwise_conv1d_silu,
@@ -25,8 +24,8 @@ NVIDIA hardware. Lower is better.
    - `cake-core/src/backends/cuda/compat/` — CUDA version compatibility layer.
    - `cake-core/src/backends/mod.rs` — ComputeBackend trait definition.
    - `cake-core/benches/bench_utils.rs` — Backend operation benchmarks (fused_*_gpu variants).
-4. **Run prepare.sh**: `bash autoresearch/backends/cuda/prepare.sh`
-5. **Confirm baseline** and start experimenting.
+3. **Run prepare.sh**: `bash autoresearch/backends/cuda/prepare.sh`
+4. **Confirm baseline** and start experimenting.
 
 **IMPORTANT**: This task requires an NVIDIA GPU with CUDA support. Build with `--features cuda`.
 

autoresearch/backends/metal/program.md

@@ -10,9 +10,8 @@ memory-bandwidth-bound at ~200-400 GiB/s depending on the chip. Lower is better.
 
 ## Setup
 
-1. **Agree on a run tag**: e.g. `mar23`. Branch: `autoresearch/metal/<tag>`.
-2. **Create the branch**: `git checkout -b autoresearch/metal/<tag>`
-3. **Read the in-scope files**:
+1. **Create the branch**: `git checkout -b autoresearch/backends/metal`
+2. **Read the in-scope files**:
    - `cake-core/src/backends/metal/ops.msl` — Metal Shading Language compute kernels.
      **Primary target.** Contains all fused ops: silu_mul, stable_softplus, rms_norm_gated,
      add_rms_norm, rms_norm_channel, add3, add_scaled, depthwise_conv1d_silu,
@@ -21,8 +20,8 @@ memory-bandwidth-bound at ~200-400 GiB/s depending on the chip. Lower is better.
      implementation, MSL compilation, command buffer management, synchronization logic.
    - `cake-core/src/backends/mod.rs` — ComputeBackend trait definition.
    - `cake-core/benches/bench_utils.rs` — Backend operation benchmarks.
-4. **Run prepare.sh**: `bash autoresearch/backends/metal/prepare.sh`
-5. **Confirm baseline** and start experimenting.
+3. **Run prepare.sh**: `bash autoresearch/backends/metal/prepare.sh`
+4. **Confirm baseline** and start experimenting.
 
 **IMPORTANT**: This task requires a Mac with Apple Silicon or Intel GPU. Build with `--features metal`.
 

autoresearch/backends/vulkan/program.md

@@ -10,9 +10,8 @@ compute shaders. Lower is better.
 
 ## Setup
 
-1. **Agree on a run tag**: e.g. `mar23`. Branch: `autoresearch/vulkan/<tag>`.
-2. **Create the branch**: `git checkout -b autoresearch/vulkan/<tag>`
-3. **Read the in-scope files**:
+1. **Create the branch**: `git checkout -b autoresearch/backends/vulkan`
+2. **Read the in-scope files**:
    - `cake-core/src/backends/vulkan/ops.wgsl` — WGSL compute shaders. **Primary target.**
      Contains: elementwise ops (silu_mul, add3, add_scaled, rms_norm_gated, add_rms_norm,
      stable_softplus, exp_mul, sub_mul) and matrix operations (GEMV with tiled 4×64 layout,
@@ -24,8 +23,8 @@ compute shaders. Lower is better.
    - `cake-core/benches/bench_vulkan.rs` — Vulkan-specific benchmarks (GEMV, GEMM, silu_mul,
      add3, rms_norm, MLP, dispatch overhead).
    - `cake-core/benches/bench_utils.rs` — Backend operation benchmarks.
-4. **Run prepare.sh**: `bash autoresearch/backends/vulkan/prepare.sh`
-5. **Confirm baseline** and start experimenting.
+3. **Run prepare.sh**: `bash autoresearch/backends/vulkan/prepare.sh`
+4. **Confirm baseline** and start experimenting.
 
 **IMPORTANT**: This task requires a Vulkan 1.3+ GPU. Build with `--features vulkan`.
 

autoresearch/kernels/attention/program.md

@@ -10,9 +10,8 @@ single most expensive operation in every transformer forward pass. Lower is bett
 
 ## Setup
 
-1. **Agree on a run tag**: e.g. `mar23`. Branch: `autoresearch/attention/<tag>`.
-2. **Create the branch**: `git checkout -b autoresearch/attention/<tag>`
-3. **Read the in-scope files**:
+1. **Create the branch**: `git checkout -b autoresearch/kernels/attention`
+2. **Read the in-scope files**:
    - `cake-core/src/models/common/attention.rs` — CausalSelfAttention: Q/K/V projection,
      rotary position embeddings (RoPE), optional QK normalization, grouped-query attention
      (GQA) head repeat, scaled dot-product attention dispatch, output projection.
@@ -27,8 +26,8 @@ single most expensive operation in every transformer forward pass. Lower is bett
    - `cake-core/benches/bench_attention.rs` — Attention benchmarks.
    - `cake-core/benches/bench_linear_attn.rs` — GatedDeltaNet linear attention benchmarks.
    - `cake-core/benches/bench_cache.rs` — Cache benchmarks.
-4. **Run prepare.sh**: `bash autoresearch/kernels/attention/prepare.sh`
-5. **Confirm baseline** and start experimenting.
+3. **Run prepare.sh**: `bash autoresearch/kernels/attention/prepare.sh`
+4. **Confirm baseline** and start experimenting.
 
 ## Files You May Modify
 

autoresearch/kernels/fused-ops/program.md

@@ -11,9 +11,8 @@ for micro-optimization. Lower is better.
 
 ## Setup
 
-1. **Agree on a run tag**: e.g. `mar23`. Branch: `autoresearch/fused-ops/<tag>`.
-2. **Create the branch**: `git checkout -b autoresearch/fused-ops/<tag>`
-3. **Read the in-scope files**:
+1. **Create the branch**: `git checkout -b autoresearch/kernels/fused-ops`
+2. **Read the in-scope files**:
    - `cake-core/src/backends/cpu/mod.rs` — CPU implementations of all fused ops. These are the
      default fallback and what benchmarks measure without GPU features.
    - `cake-core/src/backends/mod.rs` — ComputeBackend trait defining the fused op interface.
@@ -23,8 +22,8 @@ for micro-optimization. Lower is better.
      `fused_rms_norm_channel_cpu`, `fused_depthwise_conv1d_silu_cpu`, `fused_depthwise_conv1d_bias_cpu`,
      and GPU variants.
    - `cake-core/benches/bench_helpers.rs` — Shared helpers.
-4. **Run prepare.sh**: `bash autoresearch/kernels/fused-ops/prepare.sh`
-5. **Confirm baseline** and start experimenting.
+3. **Run prepare.sh**: `bash autoresearch/kernels/fused-ops/prepare.sh`
+4. **Confirm baseline** and start experimenting.
 
 ## Files You May Modify
 

autoresearch/kernels/quantization/program.md

@@ -10,9 +10,8 @@ for quantized models, making it one of the highest-frequency operations. Lower i
 
 ## Setup
 
-1. **Agree on a run tag**: e.g. `mar23`. Branch: `autoresearch/quantization/<tag>`.
-2. **Create the branch**: `git checkout -b autoresearch/quantization/<tag>`
-3. **Read the in-scope files**:
+1. **Create the branch**: `git checkout -b autoresearch/kernels/quantization`
+2. **Read the in-scope files**:
    - `cake-core/src/utils/quantization.rs` — GPTQ 4-bit dequantization, FP8 blockwise
      dequantization, weight format detection, quantization utilities.
    - `cake-core/src/backends/cpu/mod.rs` — CPU FP8 dequant: software decoding of E4M3
@@ -26,8 +25,8 @@ for quantized models, making it one of the highest-frequency operations. Lower i
      FP8 blockwise).
    - `cake-core/benches/bench_flux.rs` — FP8 linear forward, F8 dequant benchmarks.
    - `cake-core/benches/bench_utils.rs` — fp8_linear_forward_cpu, fp8_linear_forward_with_bias_cpu.
-4. **Run prepare.sh**: `bash autoresearch/kernels/quantization/prepare.sh`
-5. **Confirm baseline** and start experimenting.
+3. **Run prepare.sh**: `bash autoresearch/kernels/quantization/prepare.sh`
+4. **Confirm baseline** and start experimenting.
 
 ## Files You May Modify
 

autoresearch/models/image-generation/program.md

@@ -9,15 +9,14 @@ text encoders (T5, CLIP), timestep embedding, and FP8 linear layers. Lower is be
 
 ## Setup
 
-1. **Agree on a run tag**: e.g. `mar23`. Branch: `autoresearch/image-generation/<tag>`.
-2. **Create the branch**: `git checkout -b autoresearch/image-generation/<tag>`
-3. **Read the in-scope files**:
+1. **Create the branch**: `git checkout -b autoresearch/models/image-generation`
+2. **Read the in-scope files**:
    - `cake-core/src/models/flux/` — FLUX transformer, T5/CLIP encoders, VAE, shardable variants.
    - `cake-core/src/models/sd/` — Stable Diffusion: CLIP, VAE, U-Net, scheduler.
    - `cake-core/benches/bench_flux.rs` — FLUX component benchmarks.
    - `cake-core/benches/bench_helpers.rs` — Shared helpers.
-4. **Run prepare.sh**: `bash autoresearch/models/image-generation/prepare.sh`
-5. **Confirm baseline** and start experimenting.
+3. **Run prepare.sh**: `bash autoresearch/models/image-generation/prepare.sh`
+4. **Confirm baseline** and start experimenting.
 
 ## Files You May Modify
 

autoresearch/models/moe/program.md

@@ -10,9 +10,8 @@ where K experts are selected from N total for each token. Lower is better.
 
 ## Setup
 
-1. **Agree on a run tag**: e.g. `mar23`. Branch: `autoresearch/moe/<tag>`.
-2. **Create the branch**: `git checkout -b autoresearch/moe/<tag>`
-3. **Read the in-scope files**:
+1. **Create the branch**: `git checkout -b autoresearch/models/moe`
+2. **Read the in-scope files**:
    - `cake-core/src/models/common/mod.rs` — MoE layer: router, top-k gating, expert dispatch, combine.
    - `cake-core/src/models/common/expert_provider.rs` — ExpertProvider trait + StackedResidentProvider + IndividualResidentProvider.
    - `cake-core/src/models/common/disk_expert_provider.rs` — DiskExpertProvider for memory-mapped expert weights.
@@ -21,8 +20,8 @@ where K experts are selected from N total for each token. Lower is better.
    - `cake-core/benches/bench_moe.rs` — MoE forward benchmarks.
    - `cake-core/benches/bench_expert_provider.rs` — Expert weight access benchmarks.
    - `cake-core/benches/bench_blocks.rs` — MoE block benchmarks (qwen3_moe_block, qwen3_5_moe_block).
-4. **Run prepare.sh**: `bash autoresearch/models/moe/prepare.sh`
-5. **Confirm baseline** and start experimenting.
+3. **Run prepare.sh**: `bash autoresearch/models/moe/prepare.sh`
+4. **Confirm baseline** and start experimenting.
 
 ## Files You May Modify
 

autoresearch/models/text-inference/program.md

@@ -12,9 +12,8 @@ with small dimensions (hidden=64) for fast iteration. Lower is better.
 
 To set up a new experiment run, work with the user to:
 
-1. **Agree on a run tag**: propose a tag based on today's date (e.g. `mar23`). The branch `autoresearch/<tag>` must not already exist.
-2. **Create the branch**: `git checkout -b autoresearch/text-inference/<tag>` from the current branch.
-3. **Read the in-scope files** for full context:
+1. **Create the branch**: `git checkout -b autoresearch/models/text-inference`
+2. **Read the in-scope files** for full context:
    - `cake-core/src/models/common/attention.rs` — CausalSelfAttention: Q/K/V projection, rotary embeddings, scaled dot-product attention, output projection.
    - `cake-core/src/models/common/mlp.rs` — MLP: gate/up projection, SiLU activation, down projection.
    - `cake-core/src/models/common/transformer.rs` — TransformerBlock: attention + MLP + residual + RmsNorm composition.
@@ -27,9 +26,9 @@ To set up a new experiment run, work with the user to:
    - `cake-core/benches/bench_blocks.rs` — Transformer block benchmarks.
    - `cake-core/benches/bench_cache.rs` — Cache benchmarks.
    - `cake-core/benches/bench_helpers.rs` — Shared benchmark helpers (test_config, make_tensor, etc.).
-4. **Run prepare.sh**: `bash autoresearch/models/text-inference/prepare.sh`
-5. **Confirm baseline**: Check that `baseline.txt` and `experiments.tsv` were created.
-6. **Start experimenting**.
+3. **Run prepare.sh**: `bash autoresearch/models/text-inference/prepare.sh`
+4. **Confirm baseline**: Check that `baseline.txt` and `experiments.tsv` were created.
+5. **Start experimenting**.
 
 ## Files You May Modify