Let a container serve both D8VK and Zink on a pre-1.2 ICD

[abesmon]

Two control points and one emulation. All off by default — with no
environment set, this changes nothing for any driver.

Motivation

Newer D8VK calls Vulkan 1.3 core entry points. On a Vulkan 1.1 ICD those function
pointers do not exist, and calling one surfaced in 32-bit Wine as an assertion in
loader_thunks.c rather than a useful Vulkan error. A container can report a
higher core version, but then the wrapper must actually lower the calls that
version implies.

The catch: raising the reported version raises it for every client, and Zink
changes ABI at Vulkan 1.2. Above it, it requests features through the
VkPhysicalDeviceVulkan12Features / Vulkan13Features aggregates instead of the
extension feature structs an older ICD understands — and process_pnext_chain
drops an aggregate the base driver is too old to have. So the moment D3D8 works,
OpenGL stops.

That is why these ship together: raising the version for D8VK is what creates the
need for a per-engine version control point, and reporting 1.1 to Zink is what
makes it ask for VK_KHR_imageless_framebuffer as an extension the driver lacks.

Technical basis

Vulkan 1.3 promoted VK_KHR_synchronization2 (vkQueueSubmit2),
VK_EXT_extended_dynamic_state and …state2, and VK_KHR_dynamic_rendering.
Vulkan 1.2 promoted VK_KHR_imageless_framebuffer.
See the promotion table.

A promoted entry point is only guaranteed where the core version is. The 1.1-era
ICD here still exposes the pre-promotion EXT entry points, so the lowering
enables the EXT names by name and routes core calls to them; where no EXT
equivalent exists (vkQueueSubmit2), it translates to the legacy operation.

Before / after

	before	after
D3D8 on a 1.1 ICD	blank window, 3 FPS, never renders	renders, 55 avg fps
OpenGL when the container raises the version for D3D8	black screen, 0 FPS	renders, 60 fps
a driver lacking vkQueueSubmit2	called through a NULL pointer	VK_ERROR_FEATURE_NOT_PRESENT
everything else	—	unchanged, measured

Why this shape

The wrapper decides nothing; the container operator does. This ends a chain of
narrowing guesses — an earlier revision inspected driverID and picked a version
itself, a later one hardcoded a zink-named variable, this one removes the
guessing:

const char *override_engine = getenv("WRAPPER_ENGINE_NAME");
if (engine && override_engine && !strcmp(engine, override_engine))

Both variables are required, so an unset knob is identical to the code being
absent. The match is an exact strcmp, not a substring, so no unrelated engine
is silently caught and no engine is named in the code — the Zink case is one
instantiation, not the mechanism.

That is also the whole platform-safety argument: there is no default to regress.
The diff was audited hunk by hunk for this, and three places that ran regardless
of the knobs were brought behind them — the pNext unlink of the imageless
feature (keyed on the driver's extension list, it would have stripped a
legitimate request on a 1.2 driver supporting the feature as core), the
command-buffer walk in ResetCommandPool, and the per-view hash insert in
CreateImageView.

Flags, including for non-NVIDIA

Nothing is driver-gated. Any driver hitting the same gaps can opt in.

variable	when to set it
WRAPPER_EMULATE_VULKAN_1_3=1	base driver below 1.3 and a client calls 1.3 core entry points (D8VK). Inert on a 1.3+ driver.
WRAPPER_ENGINE_NAME + WRAPPER_ENGINE_VK_VERSION	one engine needs a different reported version than the rest. Both required.
WRAPPER_EMULATE_IMAGELESS_FRAMEBUFFER=1	a client asks for VK_KHR_imageless_framebuffer and the driver lacks it.

The Tegra configuration, as a worked example:

WRAPPER_VK_VERSION=1.3.178          # what D8VK needs
WRAPPER_ENGINE_NAME=mesa\ zink      # put Zink back on the older ABI
WRAPPER_ENGINE_VK_VERSION=1.1.178
WRAPPER_EMULATE_VULKAN_1_3=1
WRAPPER_EMULATE_IMAGELESS_FRAMEBUFFER=1

Note the escaped space. WRAPPER_ENGINE_NAME is an exact match and Zink's
engine name is mesa zink; in a launcher whose environment field is
space-separated, writing it unescaped silently yields
WRAPPER_ENGINE_NAME=mesa and the override never fires. Each control point logs
when it engages, which is how to confirm it took:

Lowering Vulkan 1.3 core calls for a 1.1 base driver (WRAPPER_EMULATE_VULKAN_1_3)
Reporting Vulkan 1.1 to engine 'mesa zink' (WRAPPER_ENGINE_VK_VERSION)
Emulating VK_KHR_imageless_framebuffer

Evidence

Switch v1, Tegra X1, NVIDIA proprietary 111.0.0 (Vulkan 1.1). Necessity was
measured by toggling one binary — an unset knob equals the commit being absent,
so this removes build variance. Removing each in turn: D3D8 drops to a blank
window at 3 FPS; Balatro goes black at 0 FPS; Balatro exits at
CreateGraphicsPipelines. For the first, diffing enabled device extensions
between arms shows exactly VK_EXT_extended_dynamic_state and …state2.

No regression elsewhere: 21-point render matrix, three passes, against the
preceding upstream build — 21/21, and D3D9/D3D10/D3D11 unmoved in both bitnesses.

Limitations

• Not general Vulkan 1.3. QueueSubmit2 lowering forwards pNext into a
  legacy VkSubmitInfo without proof that sync2-only chains are filtered or
  absent; dynamic-rendering lowering has no full depth/stencil resolve; colour
  attachments are capped at eight. It is a shim for the calls D8VK makes.
• One driver was available. Everything above is Tegra X1 — which is exactly
  why nothing is on by default.