GLU image scaling contract

gluScaleImage(format, widthIn, heightIn, typeIn, dataIn, widthOut, heightOut, typeOut, dataOut) is API id 4594, GL command opcode 141, with nine physical stdcall argument dwords (40-byte cleanup including the return address). Both identifiers are appended. Existing API rows and GL opcodes retain their original order.

The generated dispatcher uses the existing $handle_gpu_api route. The WAT encoder captures the stack and flushes opcode 141 synchronously; production OpenGLHostBridge.replay returns its GLU status and writes guest memory before the guest resumes. Input/output pixels are accessed through the canonical guest translator while the producer is parked. The operation executes immediately during display-list compilation, without recording a list command.

The implementation supports the GLU reference's color/index/depth formats: COLOR_INDEX, STENCIL_INDEX, DEPTH_COMPONENT, RED, GREEN, BLUE, ALPHA, RGB, RGBA, LUMINANCE, LUMINANCE_ALPHA, BGR and BGRA. Input and output independently accept BYTE, UNSIGNED_BYTE, SHORT, UNSIGNED_SHORT, INT, UNSIGNED_INT, FLOAT, BITMAP and the twelve RGB/RGBA packed pixel types (332/233_REV, 565/565_REV, 4444/4444_REV, 5551/1555_REV, 8888/8888_REV and 1010102/2101010_REV). Three-component packed types require RGB; four-component packed types require RGBA or BGRA. No valid type is accepted as an unimplemented success stub.

Filtering uses a box footprint for reduction and linear interpolation for enlargement, separately per axis. Edges wrap as in the SGI GLU implementation. Conversion uses the SGI-style unsigned 16-bit intermediate, including its limited precision and legacy signed/index conversions. This is not a claim of bit-identical results to every native GLU implementation. FLOAT color values in [0, 1] are quantized to that intermediate; arbitrary nonfinite float behavior is not qualified.

PACK and UNPACK alignment, row length, skipped rows/pixels, byte swapping and bitmap LSB-first modes are per-context state. Client pixel-store push/pop preserves them; integer queries expose them. Row padding and unrelated bitmap bits remain untouched. Input is snapshotted before output, including when buffers overlap. Sparse pages may have nonadjacent physical backing.

Return codes are 0 (success), 100900 (GLU_INVALID_ENUM), 100901 (GLU_INVALID_VALUE), 100902 (GLU_OUT_OF_MEMORY), and 100904 (GLU_INVALID_OPERATION for incompatible packed formats). A zero dimension is a no-op, with the SGI zero-before-negative validation order. Negative dimensions otherwise fail. Scratch storage is bounded to a combined 8,388,608 16-bit input/output components (16 MiB); exceeding it returns OUT_OF_MEMORY. Allocation RangeError also returns OUT_OF_MEMORY.

As an emulator safety extension, null, unmapped, overflowing or out-of-linear- memory guest spans return INVALID_VALUE. Every touched page of both buffers is checked before writes. There is no partial output on those validation failures. Concurrent guest remapping during a call is not qualified. The existing current-context requirement of the GL bridge applies.

Run node test/test-glu-scale-image.js for production bridge and replay tests, hash lookup/dispatch ABI checks, stable historical order, identity and resize filters, pixel storage, scalar/packed/bitmap conversion, sparse pages, overlap, errors and bounds. GLU_CONTROL_SOURCE may point to a pinned pre-change lib/gl-compat.js; the same regression must reject its missing opcode. Check generated files with tools/gen_api_table.js --check, tools/gen_dispatch.js --check, tools/gen-gl-encoder-abi.js --check and tools/check-gl-barriers.js.

References consulted:

This task verifies source generation and executable host behavior only. It does not qualify a full WAT build, original TuxRacer PE startup, rendering or gameplay, and uses no original game media or guest runtime.