9413f5cf19
The drift page said an event loses its frame and every later one. Forcing each event's one bucket on the sink's route, read against that fabric's own run, shows that holds for the two events that lose the coder's state, at 7.3 and 7.4 dB, and not for the one that costs a single symbol: read alone it moves no sample by more than 1, stays at or above 64.4 dB, and never leaves the stream. Sections 4.5 and 5 now say so, with every figure marked into readings/gop/events.json. dcvc-drift gains --force and an event subcommand, and record refuses a forced run. record no longer needs the encoder's reference: the baseline's frames, rounded half to even, must carry the digests in fixture.json. The webgpu build gets a search path to the provider's library. verify.py holds each latent to its symbols, so an event erased with its bucket is refused, and holds a forced run to the event it repeats and to the run it is read against. Section 7 states which fields it recomputes, which it takes from a reading and which it can only hold together. The change note now says what ran: only the two sink's-route fabrics went through decode_envelope; the other five ran in the loop alone. Co-Authored-By: Claude Opus 5.5 <[email protected]>
13 lines
571 B
Rust
13 lines
571 B
Rust
//! The WebGPU provider's library is a shared one, which `ort/copy-dylibs` links beside the
|
|
//! binary. A binary looks there only if told to, so this gives it that search path.
|
|
fn main() {
|
|
println!("cargo:rerun-if-changed=build.rs");
|
|
if std::env::var_os("CARGO_FEATURE_WEBGPU").is_none() {
|
|
return;
|
|
}
|
|
match std::env::var("CARGO_CFG_TARGET_OS").as_deref() {
|
|
Ok("macos") => println!("cargo:rustc-link-arg-bins=-Wl,-rpath,@loader_path"),
|
|
Ok("linux") => println!("cargo:rustc-link-arg-bins=-Wl,-rpath,$ORIGIN"),
|
|
_ => {}
|
|
}
|
|
}
|