* fix(binary-fetcher): retry zip downloads that drop mid-body
The zip path of the binary fetcher (Node.js runtimes on Windows) read the
response body outside any retry. fetchFromRegistry retries only until the
response headers arrive, so a TLS connection that closed during the body
rejected the whole install with undici's bare "terminated" error, which
names no URL.
The download and checksum step now runs under the same retry policy as
tarball downloads: 401, 403, 404 and non-retryable network errors fail at
once, anything else is retried with the configured backoff and logged
through the request-retry logger. A download that still fails names its
URL. The response status is now checked, so an error page is no longer
hashed as the archive.
The tarball downloader already retries a mid-body drop, which was checked
against a server that destroys the socket after a partial body. pnpm v12
already retries zip downloads as a whole.
Closes https://github.com/pnpm/pnpm/issues/17411
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(binary-fetcher): keep retry defaults and hide URL secrets in retries
The install client passes every retry field, with unset ones as
undefined, and `@zkochan/retry` copies undefined over its defaults, so an
unset fetch-retries disabled zip retries. Undefined fields are now
dropped before merging with the defaults.
The request-retry event no longer carries the error message, which for a
checksum mismatch embedded the raw URL, signed-URL tokens included. The
checksum error itself now shows the redacted URL.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* style(binary-fetcher): reword a comment for the spell checker
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
