Streamed file input, stencil/depth bias, blend constant, occlusion queries, and an interpreter stack fix for Halo

[replghost]

Draft: runtime support needed to run the Halo CE engine (polkavm-app-kit apps/halo) on any Host. Everything is additive within application runtime ABI 1.

Changes

Runtime file handlers and streamed file input (ADR 0003; docs/runtime/polkavm-app-abi-v1.md)

• Guests register file handlers at runtime, and a selected file can be delivered as a stream. Guests read it in bounded ranges (host_file_read) and keep a private per-selection cache (host_file_cache_*).
• Implemented in the native and browser backends. Large inputs are not copied into guest memory or packaged as assets.

Stencil and depth bias, gated by RasterFeatures bit 1 (GPU_RASTER_FEATURE_STENCIL_DEPTH_BIAS)

• Texture format 8 (Depth24PlusStencil8).
• Render-pass stencil load, store and clear (pass flags 16, 32 and 64; flag 64 adds a 4-byte stencil clear value).
• A pipeline trailer (flag 2) with stencil state and depth bias.
• Opcode 30, SetStencilReference.
• The wire decoder now rejects unknown pass and pipeline flag bits, because payload length depends on them.

Blend constant, gated by RasterFeatures bit 2 (value 4)

• Opcode 31, SetBlendConstant: 16 bytes, RGBA f32 little-endian, every value finite. Valid only inside a render pass; it applies to the draws recorded after it.
• Every pass starts with the constant 0,0,0,0, as in WebGPU. Before this change, native wgpu panicked at draw time when a pipeline used constant blend factors without a constant.

Occlusion queries, gated by RasterFeatures bit 3 (value 8)

• BeginRenderPass flag 128 declares a pass's query count and a guest token. Passes in one batch declare at most 4,096 queries in total.
• BeginOcclusionQuery (32) and EndOcclusionQuery (33) follow WebGPU's nesting rules on both Hosts.
• Each declaring pass is resolved and read back without delaying the batch. Its results arrive as GPU event 9 (token, count, u64 sample counts), in submission order.
• Native Hosts receive outstanding results from NativeGpuRenderer::poll_events.
• Native and browser Hosts now advertise RasterFeatures 15.

Interpreter stack-residency fix

• PolkaVM 0.37 (c160c13) treated a grown stack buffer's whole power-of-two capacity as resident guest stack. For a non-power-of-two declared stack, that exposed the lower guard page and corrupted Host reads of stack-backed memory.
• The engine crate is vendored under rust/vendor/polkavm with residency bounded to the declared size, and with its provenance and licences recorded. Upstream PolkaVM is unchanged.

Merge of main (#57, mediated input runtime API)

• Conflicts are resolved in the merge commit.
• A file delivered while the runtime is paused is held until resume. This is covered by a new test.

Verification

• cargo +nightly fmt --check, clippy with -D warnings, and cargo test --workspace --all-features: 193 passed.
• npm test: 216 passed. Browser assets were regenerated, and verify-browser-assets passes.
• polkavm-app-kit's Halo guests on this runtime:
  • The real engine starts on an original, freely licensed test room, on native compiler and interpreter.
  • A pixel-exact D3D frame check passes on native compiler and interpreter: 85 rectangles, 0 mismatching pixels. It covers stencil, z-bias, constant blending and occlusion counts (0 behind an occluder, exact counts in front).
  • The play check passes on native and in headless Chromium, both backends: real input moves, turns and collides with the walls.

Notes for reviewers

• The vendored engine is a local patch, not an upstream PolkaVM release. It should be dropped once an upstream PolkaVM revision includes the fix.
• No retail game data is involved in any test.