GL1.1 pixel maps (2026-11-02)

Run scratch/runs/20261102-gl-pixel-map-fresh/. Direct and queued resolved-guest fixtures each pass 935 checks. The focused suites and first locked build passed before removing exactly these six availability rejections; 64 mandatory MDK2 names remain unsupported.

Primary contracts: GL1.1 specification, sections 3.6.2/table 3.3, 3.6.4, 3.8, 5.4, 6/table 6.16; Windows setter contract; Windows getter contract.

Ten per-context tables start with one zero entry. Mapping is initially disabled, giving identity transfer. Maximum table size is 256; the six index-addressed maps require a power of two. Component maps clamp float entries and normalize unsigned entries. Color-index entries retain fractions and unsigned integer values; stencil entries round to the nearest integer. Unsigned getters normalize [0,1] to their output width; values outside that interval are unspecified by GL1.1.

RGBA transfer scales and biases before component table lookup. Color-index texture upload instead shifts/adds numeric indices and independently masks four index-to-component table addresses, regardless of MAP_COLOR. The renderer remains RGBA; storing/querying I_TO_I and S_TO_S does not add color-index framebuffer or stencil pixel consumers. Existing unsupported exports stay unsupported.

Tables and sizes are excluded from attribute stacks (table 6.16). PIXEL_MODE_BIT restores transfer enables/scales. List setters capture client buffers but modify server tables only on execution; getters execute immediately. Queued commands own setter bytes.

Native table storage preserves unsigned index words in doubles; canonical index upload slots retain numeric indices until execution. Index transfer uses 32 integer bits and eight fractional bits, with zero-fill shifts and independent table masks. Context teardown frees the tables. Native query output uses exact 2-byte or 4-byte elements after complete sparse-span validation.

The fixture resolves all six real setters/getters through guest imports and WGL, checks stdcall stack/register canaries, every target and input conversion, all default/size queries, maximum and power-of-two bounds, unchanged state/output on errors and sparse faults, context isolation, attribute exclusions, list capture/execution and immediate queries. Literal mapped pixels cover typed component/index maps, 1D/2D uploads/subimages, framebuffer reads, all four copy commands, numeric signed/unsigned/float/bitmap indices, byte/bit order, and queued input ownership. Native texture readback and the command consumer are compared separately. glGetTexImage packs stored components without applying transfer again (section 6.1.4).

Thirteen targeted semantic controls reject altered defaults, normalization, output width, lookup rounding, native/host component and index consumers, stencil rounding, power bounds/state atomicity, sparse output atomicity, list capture and queued capture. Two compiled saved-source controls reject missing genuine setters/getters. Initial fixture cleanup used the wrong worker shutdown method; it was corrected to the existing stop methods. A power-rule control originally encountered a later pointer fault because the host rejection masked the injected native error; an immediate native size assertion now proves atomicity before host error retrieval. Production semantics did not change for these fixture corrections.

Validation: 935 direct and 935 queued resolved stdcall checks; 13 semantic mutations and two saved-old-source controls; 20 focused commands covering boolean/pointer/double/texture queries, 1D textures, texture copies, scalar readback, unpack alignment, attributes, lists, native/JS encoders, worker execution, isolated typed dispatch, fixed function rendering, drawable color buffers and sparse arrays. Availability and the final locked full build verify the final source. Source hashes and test logs are in the ignored run directory.

Only the six pixel-map API rows lose gl_unimplemented; every other row, API identity and hash is preserved. Their WGL and static resolver positive checks replace exactly six mandatory rejections, reducing 70 to 64. No original MDK2 replay, download, subworker, public deployment or held GTA/D3D work was performed.