Download artifacts as parallel byte ranges that resume after a dropped connection

[joshuaswarren]

On an M2 Max, payload part00 (1.99 GB) came down at 2.9 MB/s on one stream for 11 minutes. Part01 then came down at 103 MB/s. Each part is one downloadTask with no Range and no resume, so one slow connection sets the pace, and a dropped one starts the part over.

RangedArtifactDownloader fetches 64 MiB ranges over four sessions and writes them in place. It resumes a dropped range from its last byte, and it falls back to the single-stream downloader when the server ignores Range. A range is accepted only as a 206 for exactly the requested bytes of an object of the pinned size. The stager's size and SHA-256 checks per part and for the whole payload are unchanged.

Live, on a Mac Studio on 2026-10-07, part01 of the team-preview payload (1,546,348,246 B) from GitHub release assets, SHA-256 checked against the catalog on every run:

	MB/s
main, single stream	74.0, 130.4
this branch, ranged	241.3, 165.2, 253.0

Three new tests run against a local HTTP/1.1 range server and pass three runs in a row:

• testRangedDownloadResumesADroppedRangeFromItsLastByte: the second request is bytes=150000-299999.
• testRangesRunInParallelOverSeparateConnectionsAndProgressReachesTheTotal: 16 ranges, at least two in flight on at least two connections, and progress ends at the size and never goes down.
• testServerThatIgnoresRangeFallsBackToOneStream.

Main's downloader fails the resume case with NSURLErrorDomain -1005, "The network connection was lost."

Checked at bb31f19, on main 27bdf16 (no new commits upstream), on a Mac Studio with macOS 26.6.2, Xcode 27.0 and Swift 6.4:

• swift-format lint --strict: exit 0.
• Debug swift test: exit 0 (UXCore 147, Lifecycle 3, TrustCore 522 with the 1 existing skip, App 8).
• Release swift test: exit 0 (136, 3, 522, 5).
• ./test/all: exit 0.

Risks:

• It opens four connections to one host instead of one. GitHub and Fastly served every range as 206 here. A publisher that limits connections per client could throttle or refuse. A refusal (a status other than 206 or 200) fails the attempt and does not fall back.
• I measured the speedup on a warm object from one host, in a handful of runs. I did not reproduce the M2's original case of one connection stuck at 2.93 MB/s.
• Resume works within one run only. An app restart still restarts an unfinished part.
• If a server ignores Range after some ranges have started, progress steps back once when the single-stream fallback starts from 0.

To go back to one stream and keep the code, set downloader = ProgressReportingArtifactDownloader() in VerifiedArtifactStager.init(). One line.