libusb: deliver hotplug ENUMERATE pass asynchronously on the event context

[Youw]

Implements the hotplug contract documented in hidapi/hidapi.h (#790) for the libusb backend: the HID_API_HOTPLUG_ENUMERATE initial pass is now delivered asynchronously on the same internal event context as live events, never from within hid_hotplug_register_callback() and never on an application thread.

What changed

• Registration-time replay snapshot. hid_hotplug_register_callback() no longer invokes the callback synchronously. When HID_API_HOTPLUG_ENUMERATE is set (and DEVICE_ARRIVED is requested), it builds a deep-copied snapshot of the matching entries of the internal device cache (hid_hotplug_context.devs) under the hotplug mutex and stores it on the new callback record (replay).
• Queue marker wake-up. A lightweight "replay marker" message (device == NULL) is enqueued on the existing callback-thread queue so the snapshot is delivered promptly even when there is no live hotplug traffic.
• Pass-before-live ordering. Before dispatching any live event to a callback with a pending snapshot, hid_internal_invoke_callbacks() flushes that callback's replay first. Combined with the snapshot being taken from the cache (which does not yet contain arrivals still sitting in the queue backlog), this preserves exactly-once per device connection and arrival-before-left for ENUMERATE registrations. A non-zero callback return during the pass stops the remainder of the pass and deregisters the callback, matching the live-event behavior.
• device->next is NULL for every invocation, including multi-interface/composite devices (each interface entry is delivered as a standalone invocation; the internal cache chain is preserved for LEFT matching). This closes the old TODO in the arrival path.
• Error strings. All failure paths of hid_hotplug_register_callback() / hid_hotplug_deregister_callback() (argument validation, missing LIBUSB_CAP_HAS_HOTPLUG, libusb_init / libusb_hotplug_register_callback failures, unknown handles) now set the global error via the hid_error implementation for libusb #708 infrastructure, and *callback_handle is zeroed on any registration failure.
• Shutdown/teardown rework. The event threads now wind down on a dedicated shutdown_pending flag instead of watching the callback list, and the thread join is deferred to the next registration or hid_exit():
  • removing the last callback from within a callback (non-zero return or in-callback deregister) can no longer self-join the event threads;
  • deregistering the last callback from an application thread while events are still queued can no longer deadlock (previously the join was performed while holding the mutex that the callback thread needs to drain its queue);
  • the message queue (device references and markers) is drained before the hotplug libusb context is destroyed, and undelivered replay snapshots are freed on deregistration and hid_exit().
• Registration is thread-safe end-to-end: concurrent registrations no longer race in the implicit hid_init() or in the machinery initialization (guarded by a statically-initialized mutex).

Verification

• gcc -std=c99 -fsyntax-only -Wall -Wextra -pedantic -Werror and g++ -x c++ -Wall -Wextra -Werror both clean; full CMake build with HIDAPI_WITH_LIBUSB=TRUE and tests passes.
• A smoke test (registration/deregistration, argument validation and error strings, ENUMERATE with no devices, no synchronous invocation from register, register/deregister churn, hid_exit() teardown with a live registration, re-init after hid_exit()) passes cleanly under AddressSanitizer/UBSan (no leaks) and ThreadSanitizer (no warnings), including a 4-thread concurrent register/deregister stress.
• Live device plug/unplug was not exercised (no USB passthrough available in the build environment).

Addresses #792 and #793 for the libusb backend.

Review round 1

Follow-up commit addressing the review findings. Each item below is covered by a regression test that fails when the fix is reverted (see Verification (round 1)).

Correctness / lost events

• An arriving device is now delivered exactly once, even to a callback registered from within another callback. The arrival path invoked the callbacks for each interface before appending the new entries to the device cache, and recomputed the "do not deliver to callbacks registered during this dispatch" boundary per interface. A callback registered from inside another callback during a live arrival therefore snapshotted a cache that did not yet contain the arriving device and was excluded from the in-flight dispatch, so the device reached it through neither path (and was later reported as LEFT without a preceding ARRIVED); for multi-interface devices the remaining interfaces were delivered twice instead. The infos are now appended to the cache before any callback runs, and the boundary is computed once per hotplug message.
• The libusb listener is armed before the initial enumeration, and arrivals are deduplicated against the cache. Previously the snapshot was taken first: a device connecting in between was in neither the snapshot nor any live event (libusb does not back-fill without LIBUSB_HOTPLUG_ENUMERATE), and its later removal was silently dropped as well. The reverse order can only report a device twice, which the arrival path now suppresses using the same identity the removal path matches on.
• A genuine hid_enumerate() failure is no longer mistaken for an empty system. The unconditional error reset masked it, so the registration succeeded with an empty cache and every already-connected device stayed invisible to that callback and all later ones. An enumeration failure now fails the registration and preserves the libusb error; only the genuinely-empty case resets it.

Thread safety

• The global error state is serialized. hid_hotplug_register_callback() / hid_hotplug_deregister_callback() are thread-safe by contract, but their failure paths wrote last_global_error outside any lock: two concurrent calls could double-free the stored error string. All mutations now go through a file-static mutex in the lowest-level writer.
• mutex_ready is no longer read without synchronization, and the hotplug mutex is no longer destroyed by hid_exit() while another thread is about to lock it. The flag is a one-way latch, written and read under the init mutex; the mutex and the thread states it advertises are reused by a later registration rather than destroyed.
• The deferred join is awaited on a condition variable instead of a busy unlock/lock spin, which was unfair and could livelock under a real-time scheduling policy (the spinner starving the joiner).

Lifecycle

• Deregistering the last callback from an application thread now winds the event threads down synchronously. It used to return while they were still winding down, so deregister + dlclose raced the threads into unmapped code, and the header's "the machinery runs until the last callback is deregistered" was only eventually true. The join is still deferred when the shutdown is initiated from the event thread itself, which cannot join itself.
• Deregistering an already-removed callback fails again instead of returning success a second time: a handle whose removal was postponed (an in-callback deregistration) no longer matches the lookup.
• Handle allocation refuses to overflow. next_handle++ at INT_MAX was signed-overflow UB, and the wrap-around to 1 could collide with a live handle; registration now fails with an error instead.

Failure paths

• The ENUMERATE snapshot is all-or-nothing: a failed device-info copy used to be skipped, silently hiding a connected device from that callback forever. It now unwinds the registration.
• A failed replay-marker enqueue used to be ignored, deferring the whole pass indefinitely when no live event followed; it now fails the registration too. (A dropped live event on enqueue failure is pre-existing and unavoidable - libusb offers no way to postpone or retry it - and is now called out in a comment.)
• Both event threads are checked at creation. hidapi_thread_create() returned void and threads_running was set regardless, so a failed pthread_create() left a garbage pthread_t to be joined later (UB) and a queue that never drained. It returns a status now, the callback thread is started by the registration (rather than by the libusb thread, which had no way to report a failure), and either failure fully unwinds the registration.

Verification (round 1)

• gcc -std=c99 -pedantic -Wall -Wextra -Werror and g++ -x c++ -Wall -Wextra -Werror clean; CMake builds with HIDAPI_WITH_LIBUSB=TRUE, HIDAPI_WITH_TESTS=ON and with HIDAPI_BUILD_AS_CXX=ON pass.
• The smoke test was rebuilt on top of an in-process libusb replacement, so arrivals, removals, multi-interface devices, enumeration/allocation/thread-creation failures and the arm-before-snapshot window can be simulated without real hardware (no USB access is available in the build environment). 18 test cases, including nested registration during a live arrival (single- and multi-interface, asserting exactly-once delivery and no LEFT-without-ARRIVED), arrival deduplication, the double-deregister contract, the synchronous wind-down, the deferred-shutdown race, OOM injection on both snapshot paths, both thread-creation failures, handle exhaustion, hid_exit() with a live registration and with a deferred wind-down, register/deregister churn, and a 4-thread register/deregister stress with concurrent device churn.
• Every fix above was validated with a negative control: the defect was re-introduced and the corresponding test observed to fail.
• 10/10 clean runs under ASan + UBSan + LSan (no leaks, including the libusb device references) and 10/10 clean under TSan. TSan caught two real defects introduced during this round - a lock-order inversion between the init mutex and the hotplug mutex (a nested registration from a callback takes them in the opposite order) and an under-synchronized write to shutdown_pending - both of which are fixed above.
• Live device plug/unplug on real hardware is still not exercised.

Assisted-by: claude-code:claude-fable-5
Assisted-by: claude-code:claude-opus-4-8