[kleedaisuki]

Purpose

Remove the artificial pool-wide iGPU task bottleneck. Every eligible content worker now owns CPU SIMD, a CUDA stream, and a private OpenCL queue/kernel/device buffers, with one bounded shared budget and one online model per worker.

Design

• Share only the compiled OpenCL program; each owner reuses its existing CPU file-I/O buffers and reserves its own device quota.
• Keep automatic discovery nonblocking, but make forced iGPU and explicit train wait for a peer's one-time probe rather than silently using CPU.
• Retire only the failing worker's accelerator; preserve CUDA peak/contention metric meanings and add iGPU-specific telemetry.
• Fit the model with observed CPU/iGPU or CUDA overlap while measuring error against the decision-time prediction; keep exact digest verification and fallback contracts.

Verification

• Windows MSVC Debug CTest 27/27, iGPU routing repeated 100 times.
• Physical Windows OpenCL: three concurrent admitted iGPU file tasks, 3 x 32 MiB, zero CPU fallbacks; complete BLAKE3 digests match CPU and CUDA scans.
• Physical three-backend train: 3 samples each CPU/CUDA/iGPU, 8 prior hits, 3 saved keys. Forced CUDA also reached 3 concurrent tasks and produced identical digests.
• Two interleaved 5,000-small-file CPU batches differed by ~2% in opposite directions; no established performance win or regression.

Limits

Admitted iGPU file tasks do not prove concurrent hardware kernel execution. CPU and iGPU still contend for shared memory; general whole-scan speedup is not claimed. CI must verify Linux/macOS/Windows and PoCL kernel paths before merge. Details: docs/igpu-worker-concurrency.md.