[siracozmen01]

Problem

On Windows, a generic HID gamepad's report kept none of its buttons when more than 64 of them were down. Parse in src/win32_hid.c asked HidP_GetUsages for the Button page's usages down with room for 64. A device with more buttons, such as a stick or button box of 128 whose toggle switches stay down, can have more than 64 down at once. Windows' parser then answers HIDP_STATUS_BUFFER_TOO_SMALL, and the report's buttons were dropped: each button kept the state of the last report that listed 64 or fewer, and a press or release in between was lost without a word to the program (#494).

Change

• The list of usages down has room for MWIN_WIN32_HID_DOWN, one usage per bit of a raw input read (MWIN_WIN32_HID_INPUT_BYTES * 8, 8192). Every usage down takes at least one bit of the report, and a device whose report does not fit in a raw input read is already turned away, so the list always holds them all. It lives in mwinWin32Hid, made with the context, so nothing is allocated on the event path; it adds 16 KiB there.
• mwinHidSetButtons (src/hid_controls.c) finds each usage down in the numbering by a binary search, as the numbering keeps the usages in order. Before, it compared each one with every numbered button, which a list of thousands would make slow on every report.
• docs/internals/gamepads.md says how the buttons down are asked for.

Verification

• test/test_win32_hid.c's stand-in for the parser now answers as Windows' does when the list is too small, and counts every usage down as Wine's does. A new 128-button stick sends buttons 1, 3 and 65 to 128 down. Under Wine (clang 20 for MinGW), the test fails without the change ("the box's buttons read with 66 of them down") and passes with it. The real parser is run only on the Windows CI runner.
• test/fuzz_hid.c hands mwinHidSetButtons up to 255 usages and checks each numbered button against the list: down exactly where its usage is listed. It ran 120 seconds (96,067 runs) with no finding.
• These build with no error and no warning (-Werror): linux-gcc-14 and web-wasi in the CI image, linux-clang-20-asan, linux-minimal, the fuzz tree, Windows through clang-20 for MinGW, and android-emulator.
• tools/check.sh changed passes, hid_controls included.

Closes #494

[siracozmen01]

Review at 17bf80a: ready. No findings. Read, not run.

Does 8192 bound what HidP_GetUsages returns?

• A report the raw input path parses is at most MWIN_WIN32_HID_INPUT_BYTES (1024) less the header, since a device whose input report does not fit is turned away.
• The Button page's usages down come from the report's main items. A variable item gives at most one usage per bit. An array item gives one usage per slot, and a slot is at least one bit.
• So no report can list more than 8 × 1024 = 8192, and MWIN_WIN32_HID_DOWN is that. The count < MWIN_WIN32_HID_DOWN ? count : ... clamp guards the call to mwinHidSetButtons anyway.

If the parser still answers HIDP_STATUS_BUFFER_TOO_SMALL: Parse skips mwinHidSetButtons for that report, so the buttons keep their last state while the axes and hats are still read. That is main's behaviour for every report past 64 down, now reachable only by a report the bound says cannot exist. Nothing reads past hid->down.

Is the 16 KiB only in the context?

• USAGE down[8192] is a field of mwinWin32Hid, which the context holds, and Parse now writes through hid->down instead of a 64-entry local. No function takes mwinWin32Hid by value. The test uses a static s_hid.
• One transient to know of: mwinWin32HidStart resets with *hid = (mwinWin32Hid){.context = context}; (win32_hid.c:54). An unoptimized build may build that compound literal on the stack and copy it. That is now 16 KiB larger, harmless on a Windows thread's 1 MB stack. If you want "never on the stack" to hold in Debug too, a memset(hid, 0, sizeof *hid); hid->context = context; does it.

The binary search relies on hid->buttons being ascending and unique. The numbering's own check (fuzz_hid: "each button once and in order") pins that. low/high stay within buttonCount (a uint8_t), and a usage not numbered sets nothing. fuzz_hid now lists up to 255 down and checks each button against the list, so the search is exercised past the old 64.