Request UBWC-compressed swapchain buffers on Adreno

[Leb-Sun]

Stacked on #2 (aimapper backend) so a build from this branch carries both fixes. #2 stays clean for possible upstreaming; this one is deliberately vendor-specific and is not.

Evidence

Paired same-session SurfaceFlinger captures on an A840, GameSpace frame generation ON / upscaling OFF, -b driver both runs:

system Adreno driver : usage 0x10000b00   12852 KiB   compressed: true
Turnip               : usage 0x00000b00   12768 KiB   compressed: false
                             ^^^^^^^^^

12768 KiB is 1216x2688x4 exactly — linear, no metadata. 12852 KiB is that plus the UBWC metadata plane. Stable across all 3 samples in both runs. Every other buffer on the device carries 0x10000000; only Turnip's swapchain buffer lacks it.

Layer state is byte-identical between the two runs — same dataspace=0x10810000, same composition=DEVICE (2). The buffer is the entire difference.

0x10000000 is Qualcomm's private "allocate UBWC" gralloc usage bit. Mesa sets no vendor usage bits anywhere today (vk_android.c knows only HW_RENDER and HW_TEXTURE), so Turnip has never asked for compression and gralloc had no reason to grant it.

Why not the standard route

VK_EXT_image_compression_control is the vendor-neutral mechanism, and vk_android.c:317 has a TODO pointing at it. It does not help: that extension exists for an application to request compression, and WinNative/DXVK never will. The blob does not wait to be asked — it compresses by default. Matching that is a driver default, not an app-facing option.

Gating

0x10000000 lives in AIDL BufferUsage's VENDOR_MASK (bits 28-31), so it means whatever each vendor decides. This build only ever contains freedreno (-Dvulkan-drivers=freedreno), so the bit can only be emitted from Turnip's own ANDROID_native_buffer entry points. Both are covered, since which one Android 16 calls is not established.

A better long-term gate is the detected gralloc vendor — the aimapper backend from #2 already knows it loaded mapper.qti.so. Worth doing if this proves out; not worth the plumbing before we know the bit helps.

Unproven

• The frame-gen hypothesis is untested. UBWC-vs-linear is the only meaningful difference left, and the bandwidth argument is coherent (interpolation reads two frames per output frame where upscaling reads one; UBWC roughly halves that). "Only remaining difference" is not "cause".
• Forcing UBWC may break formats that cannot be compressed. Gralloc should reject or ignore rather than misrender, but that is not verified on this device. Normal gameplay with all GameSpace features off is the check.

3 lines of code; the rest of the diff is comment. Patch is idempotent and aborts rather than writing a broken tree if the #define anchor drifts while a use still applies.