Pushed a follow-up that replaces the new test's direct Path::exists() calls with sys.fs_exists(...), fixing the clippy disallowed_methods lint while keeping the regression assertions unchanged.

Pushed a CI setup fix for the failing test unit release linux-x86_64 job. The job was failing before tests because the runner's preinstalled Azure CLI apt source returned 403 during apt-get update; the workflow now disables that unrelated source before installing LLVM packages.

Review

The core fix looks correct and well-tested. normalize_bin_name() is applied consistently at all three relevant points — collision detection in add(), within-run dedup in for_each_entry(), and the actual symlink/Windows-shim setup in set_up_bin_entry() — so the normalized name flows through everywhere it needs to (including warn_missing_entrypoint). The repro case (@eslint/config-inspector → config-inspector) resolves to the same name pnpm produces, and the String bin branch stays consistent since it already basenames via default_bin_name().

Test coverage is strong: the unix test asserts both that .bin/config-inspector is created and that no .bin/@eslint subdir leaks, which directly pins the bug.

A few notes:

• Scope: please split out the CI change. This PR also bundles an unrelated CI workflow change (ci.ts + the regenerated blocks in ci.generated.yml disabling an azure-cli apt source that returns 403). That has nothing to do with npm bin names and should be its own PR to keep history focused.

• Minor — edge cases. normalize_bin_name returns "" for inputs like "" / "/", and passes bare ".."/"." through unchanged (embedded separators are stripped, so "../foo" → "foo", but ".." alone survives). Not a regression — these were already mishandled and stripping separators actually reduces the path-escape surface — but an early guard or a comment noting these are out of scope would make intent explicit.

• Nit. The intermediate binding in normalize_bin_name can be dropped (rename the trimmed binding and unwrap_or(trimmed)), purely cosmetic.

Net positive on security too: taking the basename closes the previous write-outside-.bin path. Approve on the substance — just please pull the CI change into a separate PR.

Pushed 036363b9208c to address this: the unrelated CI workflow workaround is removed from this PR, and normalize_bin_name() now uses the trimmed binding directly with a short comment clarifying that empty/special-name validation is outside this basename normalization.