fix(encode/nvenc): cursor-blend PTX must carry a driver-portable ISA version

The vendored cursor_blend.ptx was regenerated with the CUDA 13.3 toolkit,
which stamps '.version 9.3' — and a driver whose JIT predates that ISA
refuses the whole module with CUDA_ERROR_UNSUPPORTED_PTX_VERSION (222),
silently killing host-side cursor compositing (on-glass: composite-mode
cursor invisible on driver 595.58, KWin leg). The kernels use only baseline
arithmetic, so hand-lower the version directive to 8.0 (CUDA 12.0), which
every supported Turing+ driver JITs. Both files now warn that an nvcc
regeneration re-stamps the version and must be re-lowered.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

crates/pf-encode/src/enc/linux/cursor_blend.ptx

@@ -6,7 +6,15 @@
 // Based on NVVM 7.0.1
 //
 
-.version 9.3
+// .version lowered from the generating toolkit's 9.3 (CUDA 13.3) by hand: a driver refuses PTX
+// whose ISA version is newer than its JIT (CUDA_ERROR_UNSUPPORTED_PTX_VERSION, 222 — driver
+// 595.58 = CUDA 13.2 rejects a 9.3 stamp), while a stamp OLDER than the body's actual syntax is
+// rejected as CUDA_ERROR_INVALID_PTX (218 — 8.0 was tried and failed against this body). 9.2 is
+// the highest ISA a CUDA-13.2 driver JITs and the lowest this 13.3-emitted body accepts.
+// TODO(portability): drivers older than the 13.2 era will refuse 9.2 too — the real fix is
+// regenerating cursor_blend.cu with an older toolkit (CUDA 12.x) so the body itself is old, then
+// stamping that toolkit's native version. Re-test on the oldest driver box after any regen.
+.version 9.2
 .target sm_75
 .address_size 64
 

crates/pf-encode/src/enc/linux/nvenc_cuda.rs

@@ -79,7 +79,11 @@ use nvidia_video_codec_sdk::sys::nvEncodeAPI as nv;
 /// Prebuilt PTX for the cursor-overlay blend kernels (cursor-as-metadata). Source is
 /// `cursor_blend.cu` beside this file; regenerate with
 /// `nvcc -ptx -arch=compute_75 cursor_blend.cu -o cursor_blend.ptx` after editing. JIT'd by the
-/// driver, so it runs on any Turing-or-newer GPU.
+/// driver, so it runs on any Turing-or-newer GPU. ⚠️ The `.version` stamp is load-bearing (see
+/// the comment in the .ptx): a driver refuses PTX with an ISA newer than its JIT (error 222),
+/// and a stamp older than the body's real syntax is INVALID_PTX (218) — so a new toolkit's
+/// output silently kills cursor compositing on older-driver boxes. Prefer regenerating with the
+/// OLDEST toolchain that compiles the .cu, and re-test on the oldest driver box.
 const CURSOR_PTX: &[u8] = include_bytes!("cursor_blend.ptx");
 
 // ---------------------------------------------------------------------------------------------