FEC runtime dispatch

Polaris pins nanors to b50735aedef780e533631704a8ed8cad67faf323 and compiles its rs.c, oblas_common.c, and oblas_lite.c directly. The former local ISA wrapper is removed. The later upstream revision includes the complete AVX-512 F/BW/DQ/VL admission check absent from the original Sunshine #5369 revision. No global -march=native or AVX build target is introduced.

The library selects its implementation when a codec is created. Its public init function is a no-op and its first table population is not synchronized, so host startup creates and releases one codec before HTTP/RTSP or stream threads start. Audio and video continue to own separate codecs. Codecs must not be shared concurrently based on upstream header comments alone.

Packetization, FEC percentages, oversized-frame policy, the 255-shard boundary, and the special eight-byte NVIDIA audio parity matrix are unchanged. Fixtures compare exact parity hashes against protected Polaris 7cdc9c90 and nanors 19f07b5. Tests cover unaligned buffer tails, data recovery with mixed losses, independent concurrent streams, supported size rejection, and the actual pinned selector with partial CPU capability sets. They do not assume that the decoder reconstructs missing parity or that arbitrary null shard entries are accepted.

A standalone benchmark is provided in tools/benchmarks/fec.cpp. Compile the baseline wrapper at 7cdc9c90 with its original release flags -O3 -ftree-vectorize -funroll-loops, using its pinned nanors dependency. For the baseline C++ benchmark, force-include a compatibility header containing <stddef.h> and both baseline rs.h and rswrapper.h inside extern "C". This selects the released wrapper function pointers, including its runtime ISA selection. Compile the new three C sources with -O3 and compile the identical benchmark with -std=c++20 -O3 for both versions. Do not link the raw baseline scalar implementation in place of the released wrapper.

On an Intel Core i9-14900K with GCC 16.2, seven interleaved runs pinned to one CPU measured the following median encode CPU times. Each run encoded 20,000 blocks; all before/after parity digests matched. Sizes are bytes per data shard.

Data + parity shards	Size	Before	After	Reduction
4 + 2, audio matrix	256	48.5 ns	32.5 ns	33.0%
20 + 1	1408	513.4 ns	360.7 ns	29.7%
200 + 10	1408	60.05 µs	46.27 µs	22.9%
200 + 55	1408	329.48 µs	254.42 µs	22.8%

Codec creation took 9–519 ns longer in these workloads (median wall time). These are standalone FEC measurements, not end-to-end streaming improvements. Raw samples, executable/source hashes, compiler identity, and baseline shim are retained with candidate evidence.

The parity vectors also match on native Apple ARM64/NEON and under QEMU scalar, SSSE3, and AVX2 CPU models. The available emulator cannot execute AVX-512; the synthetic capability test validates its admission rules only. Native Linux tests and the affected FEC tests under ASan/UBSan and TSan pass. Supported package builds remain part of combined candidate acceptance.