[kean]

A Nuke 13 willCache(data:image:for:pipeline:completion:) compiles against Nuke 14 with no diagnostic and was never called, so a delegate that kept private images off the disk silently started writing them. The method is a requirement again and the default async willCache forwards to it. It's soft-deprecated in the docs only: @available(*, deprecated) warns inside the default implementation and says nothing to the delegate. The guide and nuke-migrate now say it keeps working, and keep the warning for the other old delegate methods.

The pipeline waits for a closure that Nuke 13 didn't wait for, so a closure that is never called resolves to nil when it's released, and only the first call counts. With a plain continuation, as first proposed, the first case hung the load and the second crashed it.

Proof

ImagePipelineDelegateCachingTests, macOS, Xcode 27.1 RC. The suite went from 12 to 21 tests:

Delegate	main	This PR
Nuke 13, completion(nil)	callCount → 0, ["…/example.jpeg": 22789 bytes] stored	1 call, nothing stored
Same, declared in an extension	not called, 22789 bytes stored	called, nothing stored
Same, encoded image (processor, .automatic)	callCount → 0, 7 bytes stored	1 call, nothing stored
Nuke 13, replaces the data	not called, original stored	replacement stored
Nuke 13, completes on another thread	not called, original stored	replacement stored
Nuke 13, never calls completion	not called, 22789 bytes stored	image delivered, nothing stored
Nuke 13, calls completion twice	not called, stored	first call wins
No willCache at all	stored	stored
Both forms	async form wins	async form wins

A Swift 5 mode client in Release, with Nuke built at deployment target macOS 15, where the actor isolation checks fire. Same delegates, each run with a real pipeline:

completion	Plain continuation	This PR
Called synchronously, on a global queue, on main	image, 0 stored	image, 0 stored
Never called	hang, leaked its continuation	image, 0 stored
Called twice	SIGTRAP, tried to resume its continuation more than once	image, 0 stored

Cost of pipeline.willCache per store, Release, 10 interleaved A/B rounds, medians:

	main	This PR
No delegate	14.6 ns	14.3 ns
Delegate with the async form	30.4 ns	29.8 ns
Delegate with no willCache	32.5 ns	98.7 ns
Full load through a stub loader, same delegate	7.81 µs	7.74 µs
Full load through a stub loader, no delegate	10.05 µs	10.53 µs

isDefaultDelegate still returns before the delegate is touched, so the no-delegate row runs no new code. The last row is noise: the ranges overlap (7.7–13.6 and 8.0–13.5 µs) at a load average of 40. The real cost is +66 ns per disk write, and only for a delegate that doesn't implement willCache.

No new compiler warnings: a clean build of Nuke and its tests gives the same warnings as main, including the conformances above. The full Nuke scheme passes on macOS (1,739 tests, 1 skipped), and so do make snippets, make lint, generate-llms.py --check, and a docbuild.

Not verified: Xcode 26 locally, because the 26.6 install here is damaged, so CI is the first compile on it. iOS and tvOS weren't run: the change has no platform code.