[cameronwebbable]

Stacks on #20. Replaces #21, which closed when its branch was renamed; it had no reviews, so nothing was lost. Rekindled side: readwiseio/rekindled#14202 (RW-47750).

Problem

#20 added the RNCPageCurlStateMachine interface and eight tests against an empty stub. RNCWebViewPageCurl.m still carried its own state across _state, _awaitingLanding, _failedBakes and _pendingSettle, and every method touched some of them. That spread is where the four review rounds on #17 kept finding ordering bugs, and no test covered any of it.

Fix

Implement the machine and finish its tests. The machine is Foundation only: events in, effects out. Its rules are the ones the controller already followed, written down in one place.

• EnterState is emitted only when the state changes. A settle that arrives in the state it already matches changes nothing.
• A landing result is ignored unless a landing was requested. A late bridge callback after a cancel cannot move the machine.
• A settle request is acted on only in awaitingSettle, and only when its token and serial match the request still outstanding.
• A settle that arrives during a bake is kept and baked next, so a resize burst ends on the latest layout.
• Three failed bakes in a row stop the loop instead of retrying forever.

apple-tests/RNCPageCurlStateMachineTests.m has 46 tests in nine groups. The last one, testEveryOrderingKeepsTheInvariants, feeds the machine every sequence of the 14 events to depth 5 and checks the invariants after each step. That is 537,824 sequences in about 2.2 s on an M4 Pro.

Move the controller onto it. RNCWebViewPageCurl.m goes from 1404 to 1332 lines. The local enum, RNCPageCurlStateName and RNCPageCurlMaxFailedBakes are gone with the five ivars they served. The controller keeps what only it can do: the bridge calls, the landing watchdog, the renderer, the input blocker and the JS events. It feeds the machine and carries out the effects it returns in apply:.

Let the curl pan wait on WebKit's holds. A touch that moved less than 8pt in its first 350ms was handed to WebKit for good by a timer, so a drag that began with a short pause never curled, and the manager's own swipe rule turned the page with no animation. The timer is gone. The curl pan now requires WebKit's long holds to fail: the selection loupe (0.5s, 8pt), the link lift (0.65s, 10pt) and the long press (0.75s, 10pt). Those fail the moment the finger moves past their allowance, so a drag curls about 17ms after the finger passes the pan's own threshold, measured on the Duo simulator with the native log. A finger that stays still for 0.5s still selects. The holds shorter than 0.3s (the tap highlight at 0.12s, image analysis at 0.1s) keep yielding to the pan as before; waiting on them too blocked every drag that paused more than 120ms, which is how that threshold was chosen.

Grab the page under the finger. A turn started mid-page moved the far corner, so the fold trailed well to the right of the finger: about 290pt behind it 45pt into a drag from the middle of the right page. The grab point is now the touch, held no closer to the spine than grabFloor of the page width (0.3), and the fold sweeps in from the outer edge to meet the finger at sweepGain (2) times the finger's speed, so no flap appears at the grab point on the first pixel. grabFloor 1 gives the old edge grab exactly. Both are tuning keys, so the default can be changed from JS.

Curl margin taps. A margin tap turned the page with no animation even with the curl on. The manager now hands a margin tap over as a tapTurn message. Once that tap's own touchEnd has brought the controller back to idle, it curls the page from rest: an ease in and out over tapTurnDuration (0.42 s), with the bend formed over tapBendInDistance (300 pt) instead of the drag's 30 pt. Through the drag's short bend-in a timed animation makes the fold race ahead of the page edge at 3.6 times its speed and then fall back; with the release's ease-out on top, the first frame moved the fold 134 pt. Measured on the Duo simulator, the largest frame-to-frame change in fold speed is now 20 pt per frame, down from 91. With nothing baked to curl onto, or the controller busy, it calls instantlyScrollPageDown/Up the way a tap always did. The controller calls nativePageCurlAnimateTaps on the manager when it is enabled, so the manager under an older binary keeps instant taps. No state machine change: a tap turn is turnBegan from idle, like a drag.

Verification

• The test scheme is green: xcodebuild test -workspace bookwise.xcworkspace -scheme react-native-webview-Unit-Tests from reading-clients/bookwise/ios on the rekindled branch.
• Grab under the finger, recorded on the Duo simulator: a mid-page drag keeps the fold within about 25pt of the finger once it has swept in (130pt and 220pt into the drag, against 226pt and 181pt with the edge grab); edge starts, starts near the spine and back turns in the spread all curl. The back turn on a single page (closed display, portrait) was traced by hand, not recorded.
• Gesture cases on the Duo simulator after the hold change: fast and medium mid-page drags curl, an edge drag curls, a tap turns, and a finger that creeps under 8pt for 0.5s gets the selection loupe and the Define sheet, not a page turn.
• Regression pass on the iPhone Duo simulator with the native log on, reading the machine's transitions: swipe turn, tap turn, cancelled slow drag, turns on the closed display, unfold mid-book (resize rebake, spine change teardown and re-enable, kept settles collapsed to the latest), and turns across chunk edges. Every one ended in idle on the expected page.

Not in this PR

• The CI job that runs the test scheme. It lives in rekindled and is green there.
• The docs/curl-spam.sh stress run and the Metal texture check from the plan.

Bridge version

RNCPageCurlInfo is a small native module that exports protocol: 2 on iOS. The app's JS turns the curl on only when it sees that value, so JS shipped to an older binary by OTA leaves its older curl alone. Bump RNCPageCurlProtocol whenever the calls between RNCWebViewPageCurl and the content frame change.

[linear-code]

RW-47750