Resize Qwen3-VL video frames during decoding to reduce peak host memory · fixes #60985

flowchart LR
    A[Decode sampled frames] --> B{Safe to resize early?}
    B -- "supported processor,<br/>no sizing overrides" --> C[Bicubic + antialias<br/>at target size]
    B -- "unsupported config or<br/>top-level / nested override" --> D[Existing full-resolution path]
    C --> E[smart_resize is a no-op:<br/>bit-identical model inputs]
    D --> E

• Resize settings are derived from the video processor config, so nothing is hard-coded.
• A review comment exposed that sizing overrides nested under videos_kwargs bypassed the guard. I fixed it and added regression tests.
• Synthetic 720p benchmark: frame-buffer memory 13.18 → 2.11 MiB. End-to-end server memory and GPU decode paths are not yet benchmarked.

A deterministic matching-engine simulator: price-time priority, integer ticks, pre-trade risk controls, exact event/OMS replay. It is offline by design, and the README lists its non-goals.

Known limit: a 1,000,000-operation differential run was interrupted by cumulative O(N²) cancellation cost under concentrated depth. This is documented in the repo and is the next thing to fix.

A runtime intelligence platform with one loop: predict → decide → optimize → verify. It forecasts resource pressure from real host telemetry and acts only inside explicit guardrails. Each intervention is measured before and after and carries rollback metadata. Simulated components are labelled as simulated.

An interactive platform for PHY/DSP experimentation: modulation, OFDM, channel modeling, equalization, and BER/SER/EVM analysis. Live

1. A claim covers only what was measured.
2. Determinism is a testing tool.
3. Optimizations ship with evidence and a path back.
4. Non-goals are documented.

_{Open-source contributions are made independently and imply no affiliation with the projects' maintainers or sponsors.}