macOS: migrate from objc to objc2

[paravozz]

Summary

• Replaces objc 0.2.7 runtime types and macros with objc2 0.6 across src/macos/ and src/gl/macos.rs. Keeps cocoa = "0.24" for framework type bindings (NSEvent, NSView, NSWindow, etc.); pure macros migration, no public-API or behavioral change.
• Drops #![allow(unexpected_cfgs)] from both files (no longer needed: objc2's macros don't expand the buggy cargo-clippy cfg, so the warning that justified it goes silent at the source).
• Resolves the unexpected_cfgs lint failure that was making CI fail under RUSTFLAGS=-D warnings. Related to macOS builds failing with "unexpected cfg condition value: cargo-clippy" #229; this PR is one of two paths discussed there (macros-only vs. full cocoa replacement) and intentionally takes the smaller-scope macros-only route to leave the cocoa replacement decision open. Supersedes ci: register cargo-clippy cfg to silence objc 0.2.7 macro warnings #231 (the build.rs workaround), same problem fixed at the source rather than masked.
• Migration also surfaces and fixes several latent bugs in keyboard.rs and view.rs that objc 0.2's lax verification accepted but Apple's actual @encode requires; details below.

What changed

Direct replacements (mechanical):

• objc::msg_send! → objc2::msg_send! (~54 sites across view.rs, window.rs, keyboard.rs, gl/macos.rs)
• objc::class! → objc2::class! (3 sites)
• objc::sel! → objc2::sel! (26 sites)
• objc::declare::ClassDecl → objc2::runtime::ClassBuilder
• objc::runtime::Object / Class / Sel / Protocol → objc2::runtime::AnyObject / AnyClass / Sel / AnyProtocol
• objc::{Encode, Encoding} → objc2::{Encode, Encoding}
• objc::runtime::objc_disposeClassPair → objc2::ffi::objc_disposeClassPair

Less-mechanical patterns required by objc2's stricter API:

• Callback signatures use *const AnyObject / *mut AnyObject instead of &AnyObject / &mut AnyObject. objc2 0.6's MethodImplementation trait impls aren't HRT-compatible, and casting an elided-lifetime extern "C" fn(&AnyObject, ...) produces a for<'a> HRT type that fails the bound. Raw pointers carry no lifetime, dodge the issue, and faithfully match the objc_msgSend ABI.
• Callbacks marked extern "C-unwind" rather than extern "C". Plain extern "C" is undefined behavior if a foreign exception unwinds through it (Itanium ABI on macOS Apple Silicon makes this concrete). objc2's own examples use extern "C-unwind" throughout.
• bool is not objc2::Encode (objc2 docs: "bool [...] cannot be safely used in custom defined methods"), so BOOL-returning callbacks return objc2::runtime::Bool and use .as_bool() at comparison sites.
• Local CgPoint / CgSize / CgRect #[repr(C)] wrappers for the geometric types used as msg_send! returns. cocoa's NSPoint / NSSize / NSRect are external types and can't implement objc2's external Encode trait (orphan rule), so the wrappers carry the impl. From conversions to/from cocoa types at the boundary.
• ClassBuilder::new, class.add_ivar, and AnyClass::get / AnyProtocol::get take &CStr (was &str in objc 0.2). CString::new(...) for the runtime-built per-instance class name; CStr::from_bytes_with_nul(b"...\0").unwrap() for compile-time-known names (avoids c"..." literals which need edition 2021+).

Latent bugs surfaced and fixed

The migration's stricter unwinding (extern "C-unwind") and stricter @encode verification (debug_assertions mode) surfaced several bugs that have been latent in master for a while. I'm fixing them inside this PR rather than as separate PRs because they're directly entangled with the migration's stricter checks.

1. -[NSEvent isARepeat] called eagerly on flagsChanged: events (keyboard.rs::process_native_event). Per Apple docs, isARepeat raises NSInternalInconsistencyException when sent to non-key events. The result was only used for event_type == NSKeyDown (&& short-circuit), but the call was unconditional. Under extern "C" + objc 0.2 the foreign exception's behavior on macOS aarch64 happened to be survivable; under extern "C-unwind" the unwind is well-defined and silently swallows the modifier-state KeyboardEvent for every Alt/Cmd/Shift press, breaking downstream modifier tracking (textbox consumers can't see Cmd+Backspace as cx.modifiers.logo()). Verified end-to-end: an instrumented trace showed flagsChanged ENTER firing then no follow-up; gating on event_type == NSKeyDown restores the modifier-press KeyboardEvent.

2. -[NSEvent keyCode] called eagerly before event-type filter (keyboard.rs::process_native_event). Same exception class as isARepeat (Apple docs: "raises an NSInternalInconsistencyException if sent to non-key events.") The function already had a _ => return None arm acknowledging that AppKit occasionally dispatches non-key events into keyDown:/keyUp:/flagsChanged: selectors (NSAppKitDefined, NSSystemDefined, sync events around Cmd-Tab and input-source switches), but keyCode() was called before reaching that match, making the fallback dead code. Fix: hoist event-type filter above keyCode().

3. NSArray::objectAtIndex(tracking_areas, 0) without count check (view.rs::update_tracking_areas). Raises NSRangeException if index >= count. The companion site in view_will_move_to_window already guards on count != 0; mirror the same guard here for defense-in-depth.

4. Three return-type vs Apple @encode mismatches surfaced by objc2's debug-mode runtime verification:

   • view.rs::create_view: registerForDraggedTypes: declared @ (id), Apple's signature returns void (v). Fix: declare ().
   • window.rs::open_parented: addSubview: declared @, Apple returns void. Fix: declare ().
   • gl/macos.rs::create: retain declared (), Apple returns instancetype (@). Fix: declare *mut AnyObject.

   objc 0.2's msg_send! does not verify return-type encodings; objc2's does (under debug_assertions), so these mismatches that have shipped silently now panic on first debug run of cargo run --example open_window and cargo run --example render_femtovg.

Test plan

• cargo build --workspace --all-targets --all-features with RUSTFLAGS=-D warnings clean
• cargo fmt --all -- --check clean
• cargo test --workspace --all-targets --all-features clean
• cargo run --example open_window opens cleanly, dispatches mouse + keyboard events, modifier-state tracking works (Alt+G correctly synthesizes © with Modifiers(ALT), Alt release synthesizes Modifiers(0x0))
• cargo run --example render_femtovg opens cleanly, OpenGL rendering through gl/macos.rs works, mouse events dispatch
• DAW integration test against an in-tree consumer using Window::open_parented with Skia/Metal: plugin opens cleanly in Ableton Live + REAPER (with and without "Send all keyboard input to plug-in"); all keyboard/mouse/drag-drop functionality verified including modifier-held shortcuts (Cmd+Backspace, Alt+Backspace, Cmd+Left, etc.); pluginval --strictness-level 10 passes
• cargo run --example open_parented panics in softbuffer/cg.rs:36:26 null pointer dereference. Confirmed unchanged from master (same panic on a clean checkout of the pre-migration code), so this is a pre-existing softbuffer-vs-baseview lifecycle issue, not a migration regression. Worth a separate issue/PR.

What's not in this PR

Intentionally minimal. cocoa is kept as the source of framework type bindings; cocoa::base::id is kept at the public surface for raw pointer crossings; Retained<T> and MainThreadMarker patterns are NOT introduced. A follow-up PR can replace cocoa with objc2-app-kit / objc2-foundation for typed framework methods, modern memory model, and current-Xcode SDK bindings, but that's a larger refactor with its own design discussion (and probably the right call to align on with @micahrj before pursuing).

Related to #229. Supersedes #231.

[micahrj]

Forgive me if my assumption is incorrect, but this looks like a largely LLM-generated PR. This is much too big and consequential a change for me to trust an LLM to do it, and properly reviewing it will take as much work or more than just making the change myself. I would also rather make a complete move from objc and cocoa to the objc2 ecosystem rather than doing it piecemeal like this. Since this change is not actually necessary to fix the failing CI builds (#233 is enough), I am going to close this PR.