Replace warpforth-runner C++ CLI with Python cuda-python + JSON protocol

[tetsuo-cpp]

Motivation

The current warpforth-runner is a single-file C++ program that:

• Must be SCP'd to the remote GPU host and compiled with nvcc before use
• Uses CLI arguments as an ad-hoc protocol (--param i64[]:1,2,3 --output-param 0 --output-count 3)
• Outputs bare CSV to stdout (1,2,3)
• Only supports reading back a single output parameter per invocation
• Has no structured error reporting — CUDA errors go to stderr as free-form text

This makes the test harness fragile (string building/parsing in conftest.py) and hard to extend.

Proposal

Replace warpforth-runner with a Python script using cuda-python that communicates via JSON on stdin/stdout.

Input (JSON on stdin)

{
  "ptx_base64": "Ly8gUFRYIHNvdXJjZQ==",
  "kernel": "main",
  "grid": [1, 1, 1],
  "block": [64, 1, 1],
  "params": [
    { "type": "i64[]", "values": [1, 2, 3, 0, 0] },
    { "type": "f64",   "value": 3.14 },
    { "type": "i64[]", "values": [0, 0, 0, 0, 0] }
  ],
  "outputs": [
    { "param": 0, "count": 3 },
    { "param": 2 }
  ]
}

• ptx_base64: PTX source as base64 (avoids JSON escaping issues with special characters in PTX)
• params: array/scalar declarations with typed values, matching the current --param semantics
• outputs: list of params to read back (supports multiple), with optional count

Output (JSON on stdout)

Success:

{
  "status": "ok",
  "outputs": [
    { "param": 0, "type": "i64[]", "values": [10, 20, 30] },
    { "param": 2, "type": "i64[]", "values": [42, 43, 44] }
  ]
}

Error:

{
  "status": "error",
  "error": "CUDA_ERROR_INVALID_PTX: the provided PTX was invalid"
}

Benefits

• No remote compilation: pip install cuda-python on the GPU host vs nvcc compiling a C++ file
• Structured protocol: JSON is self-describing, validatable, and trivially extensible
• Multiple outputs: read back any number of params in one invocation
• Inline PTX: the kernel source is embedded in the JSON payload — no separate file upload
• Better errors: structured status/error fields instead of parsing stderr
• Simpler test harness: conftest.py builds/parses dicts instead of string-building CLI args and parsing CSV

Migration

1. Add warpforth-runner.py as a new tool (could live in gpu_test/ or tools/)
2. Update conftest.py KernelRunner to construct JSON and parse JSON responses
3. Remove the nvcc compilation step from the remote setup
4. Drop tools/warpforth-runner/ (the C++ version) once the Python version is proven
5. Add cuda-python to the project's Python dependencies