[botandbang]

What and why

The upload-cancel test still failed under load after #278. Its server read 8 KiB every 10 ms, and the test gave the server 10 s to see the connection close. The host stops sending at its next chunk once the cancel reaches it. But what it sent before the cancel still sits in the socket buffers (0.7 to 1 MB here, up to several MiB with macOS autotuning), and at that read rate draining it took 1.4 s on an idle machine and more than 10 s under load. A 15x slower server reproduces the timeout.

Reading at full speed isn't enough on its own: the cancel goes from the host to the plugin and back to the host's net/fetch before the upload sees it, and under load a fast server takes the whole 32 MiB before then.

Now the server reads once, cancels, and reads nothing more until publishing has returned. Until then the host can only fill the buffers. Then it reads the rest at once, stopping at EOF or the full size, and the test joins it for the byte count. That works because the plugin passes the cancel on to its net/fetch before it answers the hook, so by the time publishing returns the upload has been told to stop. The 10 s eventually, the per-read sleep, and the two atomics are gone.

Fixes LOC-108

How it was tested

• cargo fmt --check, and cargo clippy -p locallibrary --all-targets -- -D warnings.
• cargo nextest run -p locallibrary --test it --retries 0 on plugins::destinations::: 11 pass.
• 50 runs in a row under 45 busy loops, and 50 more beside the other 196 plugins:: tests at --test-threads 8 (load 60 to 112): all passed, with at most 680 KB of the 32 MiB read. The 196 passed too.
• With the upload's cancel check removed, it fails in 0.5 s saying all 32 MiB were sent. With a 100 ms sleep injected into the host's Sending::read, it passes in 1.4 s.
• Not run: the full suite and Windows clippy.

Checklist

• Tests cover the change

[linear-code]

LOC-108

[github-actions]

Rust coverage

From run 562. Both are held to the thresholds in .cargo/config.toml.

Tests	Lines	Functions	Regions
All	97.46% (81787/83922)	96.57% (13034/13497)	95.76% (179825/187793)
Hegel property tests alone	95.2% (79893/83922)	93.75% (12653/13497)	93.3% (175216/187793)