fix(core): a transfer cancelled before it starts spawns nothing, so it never connects

[RutaTang]

Summary

lsp::store::tests::downloads_are_https_only_and_cancellable went red on main's CI run 55 ("a cancelled download connected") and passed on the re-run. The race is the transport's, not the test's:

• net::transfer_on spawned the download thread first and looked at cancel only in the loop after it.
• The thread watches stop, which Abandon raises as the function returns — after that look — so with the flag already up, a thread scheduled ahead of its caller was past its own check and on the wire before stop was set.
• Under CPU load the store test failed 64 times in 4583 runs locally.

Fix: transfer_on reads cancel before spawning. A transfer already cancelled starts no thread and makes no connection, which is what fetch's "no connection is even attempted" promise rests on.

Regression test: net::tests::a_transfer_cancelled_before_it_starts_never_connects repeats a cancelled-at-entry transfer 200× against a listener that must see nothing. Without the fix it fails 101 runs in 137 under load; with it, neither test failed in 616.

Test plan

• cargo test -p clew-core --lib — 651 passed (git 2.55 on PATH, matching the runners)
• cargo clippy -p clew-core --all-targets -- -D warnings
• cargo fmt --all -- --check, rustdoc with -D warnings --document-private-items
• stress: fix disabled → new test fails 101/137; fix enabled → 0/616
• CI green on this PR

🤖 Generated with Claude Code

https://claude.ai/code/session_01WRJ4sDknvQJ86tFvGydgpf

---

Generated by Claude Code