refactor(ios)!: port the React Native bridge module to Swift

[kherembourg]

Phase 1 and phase 2 of the iOS Objective-C to Swift bridge migration, MOB-466.

Spec: docs/superpowers/specs/2026-09-08-ios-bridge-swift-migration-design.md
Plan: docs/superpowers/plans/2026-09-08-ios-bridge-swift-migration.md (16 tasks; the
"Orchestrator amendments" block at the top records every decision taken during execution)

2086 lines of Objective-C become six Swift files plus a 63-line export shim. No JavaScript API
change: same module name, same 63 method names, same 60 constants with the same values, same 11
events, same dictionary keys and the same per-field absence policy. Nothing under src/ is touched.

Why the shim still exists

React Native discovers a legacy module through class methods that only RCT_EXTERN_REMAP_MODULE and
RCT_EXTERN_METHOD generate, and Swift cannot emit them. So PurchaselyRN.m survives as
RCT_EXTERN_REMAP_MODULE plus 63 RCT_EXTERN_METHOD lines and the PLYRNLogWarn wrapper
(RCTLogWarn is a variadic macro Swift cannot call).

That shim is parsed as text at registration, not linked. A Swift @objc signature that disagrees
with its RCT_EXTERN_METHOD line produces no compile error and fails at run time in a client app with
"method not found". No build job catches it, so three gates do:

Gate	What it locks
BridgeExportContractTests	the 63 JS names, the module name, the 63 exported selectors, the 60 constants with their values, the 11 events
BridgeSelectorResolutionTests	every one of the 63 shim selectors actually resolves on the class
SerializationContractTests	each serializer's key set, value types and per-field absence policy

Both contract gates were written against the Objective-C output before any code moved, and each was
shown failing before it counted — on a removed export, on a renamed selector, on a changed argument
count, and on an emptied export table.

Two selectors change; neither is JS-visible

RCT_EXTERN_REMAP_METHOD is not public in RN 0.86 (RCTBridgeModule.h:310-323), so the shim cannot
remap. restoreAllProducts and silentRestoreAllProducts therefore take selectors whose first segment
is their JS name:

resolve:reject:                    ->  restoreAllProducts:reject:
silentRestoreWithResolve:reject:   ->  silentRestoreAllProducts:reject:

The other 61 were verified byte-for-byte against their Swift @objc(...) annotations, mechanically and
without sampling.

PLYSubscription stays Objective-C, on purpose

This is the one deliberate departure from the plan. PLYSubscription.init(from:) resolves .product
through ProductRepository.shared, which is internal with no injection seam — the SDK's own source
carries #warning("This decoding depends on ProductRepository, which cannot be injected"). No fixture
can be built, and integration_test/ never calls userSubscriptions either, so a port would have been
the only change here with no automated coverage at any level — on the one serializer whose absence
policy types.ts:138-141 documents to clients.

Classes/Hybrid/PLYSubscription+Hybrid.{h,m} therefore remains, reaching the Swift serializers through
a forward declaration. Because an Objective-C message send to a Swift @objc extension method is
runtime dispatch with no link-time reference, @objc on PLYPlan.asDictionary and
PLYProduct.asDictionary is permanent, and testPlanAndProductRespondToAsDictionarySelector is the
only thing that catches its removal. That was verified by mutation: dropping @objc leaves the build
green and turns that test red.

Follow-up: port it once the SDK gives ProductRepository an injection seam, which its own TODO already
anticipates.

One intentional behaviour change

The hex colour parser is now uniformly stricter on malformed input. The Objective-C used NSScanner,
which returned a colour — often opaque black — for input like "FF00ZZ", "0xFFFF" or "FF 000";
UInt32(_:radix:) returns nil, and a hex-digit guard makes the leading-sign case ("+12345") behave
the same way as the rest instead of being an exception. The path is client-reachable
(map["backgroundColors"]["light"] comes from a JS argument), so a malformed colour now falls back
instead of rendering black.

Everything else is behaviour-preserving, including two things the port had changed and which were put
back: the five array attribute setters (the Objective-C coerced elements in two of them and rejected the
whole array in the other three — user attributes drive audience targeting, so a silently partial array
shows the wrong paywall), and a nil requestId, which omits its key from the event body rather than
sending an empty string.

Also deleted

PurchaselyRN.h, PLYTransitionFactory.swift (its only reason was that Objective-C cannot construct a
PLYTransition), the stale packages/purchasely/ios/Purchasely.xcodeproj that no CI job reads, the
seven +Hybrid serializer pairs, and three dead members (sharedViewController, shouldReopenPaywall,
presentedPresentationViewController).

Verification

Check	Result
iOS unit tests	219 passed, 0 failures (83 before this work)
Example app build	BUILD SUCCEEDED
Framework layout, USE_FRAMEWORKS=static	BUILD SUCCEEDED
JS tests	283 passed, 6 suites
yarn lint / yarn typecheck	0 errors

e2e-ios T1-T30 is the real acceptance criterion — Jest mocks the native module, so it proves nothing
about a text-parsed shim reaching JS.

Release vehicle 6.2.0, not a 6.1.x patch: no client-visible feature, but the whole native layer of
one platform changes.