GL bitmap boundary contract and pending glLogicOp semantics (2026-11-02)

glLogicOp remains unsupported. The later bounded bitmap investigation below repairs demonstrated import/publication loss; it makes no logic-op or gameplay claim. Earlier blocked execution is retained as historical evidence.

Stored-width prerequisite and source-precision blocker

The subsequent run is scratch/runs/20261102-gl-logic-op-stored-width-ordering/. It retains a bounded storage repair. It does not implement or advertise glLogicOp; the sixteen-operation supported-route contract remains unpassed. The source-precision collision below is a concrete architectural blocker for a faithful downstream-only implementation.

Actual stored-width measurements and repair

Actual Chrome 151 / ANGLE Intel UHD 620 / Mesa 23.2.1 WebGL1 and WebGL2 reported 8/8/8/8 attachment bits for RGB555 bitmap drawables on unchanged production source. A packed float-color point with RGB 7/255 read [7,7,7,255] immediately, then [8,8,8,255] after publication and reimport. Twelve threshold/extreme controls distinguished private attachment values from canonical five-bit storage. Two overlapping half-alpha white points over black read 128 then 192; publication changed 192 to 189. These are ordinary float-source storage observations, not integer-color conversion or logic-operation acceptance.

Separately allocated low-bit FBOs actually reported 5/5/5/1 and completed on both WebGL versions. The repair selects that storage for canonical RGB555 bitmap masks, validates its actual precision, and preserves the format on resize. A failed precision/allocation check frees the candidate and retains the old drawable. RGBA8 remains the window/32-bit format. Other 16-bit mask layouts retain their previous backing behavior; no new arbitrary-mask support is claimed.

The renderer transport now carries the canonical channel masks, rather than assuming bpp alone identifies the format. The actual browser-worker bitmap route also selects five-bit storage. The private A1 channel is opaque and unexposed: GL alpha width queries return zero, alpha-only writes do not alter it, and canonical DIB X bits remain independent preserved padding. Storage publication no longer changes readback. The overlapping blend controls now read 132 then 197, with 197 retained after publication, for separate commands and repeated points in one packed batch. All sixteen color masks have literal pixel controls; canonical high bits and non-pixel row padding are checked.

Accumulation RETURN required a matching packed 5/5/5/1 upload on WebGL1; byte upload produced INVALID_OPERATION. WebGL2 accepted byte upload but exposed private alpha zero. The repaired RETURN keeps private alpha opaque and uses the low-bit upload type. The literal [.25,.5,.75,0] return reads [66,132,189,255] on RGB555, with no GL error; RGBA8 remains [64,128,191,0]. A second masked RETURN at .0162 distinguishes direct five-bit conversion ([8,0,8,255]) from an incorrect intermediate byte rounding ([0,0,0,255]); RGBA8 returns [4,0,4,0]. Both direct and actual browser-worker controls exercise this boundary. LOAD/ACCUM likewise reconstruct the five-bit level before normalization; LOAD of level 2 scaled by .252 then RETURN gives [8,8,8,255], rejecting byte-normalized [0,0,0,255]. This is storage behavior, not proof of the missing incoming integer-color conversion rule.

Information lost before any destination-dependent executor

OpenGL 1.1 section 2.13.9 requires unsigned integer colors to preserve their specified most significant bits when lighting is disabled and the vertex color was not clipped. A production-handler fixture and real resolved native stdcall calls measured:

Stored RGB width Unsigned glColor4ui RGB inputs Canonical float32 RGB, both inputs Required stored source values
5 0x07ffffff, 0x08000000 0.03125 0, 1
8 0x7fffffff, 0x80000000 0.5 127, 128

With lighting off, unmodified point colors, and dithering disabled, the native RGB555 stores were 0x8421 for both first-row inputs (the high bit was preexisting X padding); RGBA8 stores were 0xff808080 for both second-row inputs. Both row orientations and direct execution were measured. The actual queued-software run reproduced the collisions with eighteen GL submissions and zero fallbacks. These observations exercise supported color/draw APIs; glLogicOp lookup remained zero throughout. They are evidence of missing precision, not successful logic-operation pixels.

$gl_color_integer_component in src/09a8b-handlers-opengl.wat demotes to f32, and $handle_glColor_integer lowers to opcode 25 (glColor4f) before list capture and vertex encoding. The packed vertex contains float32 colors. Thus the two distinguishable unsigned inputs already have identical downstream state and records. No choice of destination snapshot, byte XOR, or downstream truth-table formula can reconstruct the missing source bit. For example, five-bit XOR with destination 19 requires 19 and 18 for the first pair; eight-bit XOR with destination 166 requires 217 and 38 for the second.

The next implementation step is to preserve the required integer source bits or equivalent exact conversion metadata across current color, list capture, attribute/context restore, vertex/array encoding and worker snapshots, until stored-width source conversion has consumed them. Float current-color queries can retain their defined float representation. This requires an explicit transport/state design, rather than enabling a destination-only shader against the existing float records. No unused partial truth-table implementation or availability change is retained in this repair.

Destination ordering measurements

A separate actual WebGL control implements the exact normalized form of stored-width destination INVERT (1-d), using an unattached snapshot. It does not invoke or implement the public GL logic-operation API. In both WebGL versions, two overlapping points using one snapshot leave five-bit destination 19 at 12; this is wrong for two inversions, whether submitted in one draw or two draws. Refreshing between primitives restores 19, and a third fresh inversion gives 12. The eight-bit literals are 166, 89, 166 and 89 respectively. Sampling the attached destination instead raises INVALID_OPERATION (1282) on both versions. The stale-snapshot and attached-feedback alternatives are rejected controls, not passing semantic routes.

Production WebGLBackend.draw still submits complete vertex/index ranges; createProgram still duplicates a single fragment color into MRT front/back outputs. An eventual executor must refresh after preceding overlapping work, read each selected destination independently, and apply depth/stencil effects once. That executor, the full sixteen truth tables, source conversion, blend suppression/state, errors, lists and state restoration still need their complete native/direct/actual queued/software/WebGL acceptance before lookup availability changes. DESIGN.md retains independent literal five/eight-bit all-operation tables; those tables alone are not runtime acceptance.

Validation and bounded failures

Validation details and exact final identities are in this run's result.json, final.txt, RECOVERY.md and final manifest. The retained storage changes have actual direct WebGL1/2 and actual browser-worker controls, resolved native bitmap regressions including actual queued software, focused regressions and a locked full build. Intended negative controls include restored private RGBA8 storage, private alpha queries, wrong RETURN upload/alpha and intermediate byte rounding, byte-normalized LOAD, dropped queue masks, and the preceding alpha/padding/publication mutations.

Two early local phases terminated fail-closed on a stale 1.5-second heartbeat; no disk-floor breach was measured. Their logs remain failures, including one interrupted retrieval after the remote browser had exited. A subsequent foreground supervisor uses a five-second liveness lease and an independently sampling child guard; both retain fsynced terminal minima and sampling gaps. No detached background watcher is used. No claim retroactively fills idle or preparation sampling gaps. Heavy operations run only on the recorded fresh, finite-TTL boat, with bounded direct hashed source streams and no local archive, original-game replay, fixture download or deployment.

The broad existing window-drawable test also reaches a failed-bind state assertion (bridge._current() is undefined after failed bind); the same TypeError occurs on unchanged source. Its actual worker route passes 26 checks, direct accumulation passes 584/588 checks on WebGL1/2, and worker accumulation passes 525 checks on both sources. The failing broad direct test is retained separately from passing storage controls. An initial precision fixture chose a noncolliding low eight-bit boundary and failed its own collision assertion; the corrected literal half-range pair above is preserved alongside that failure. Neither failure is counted as a passing control. A reused scratch lease filename also rejected a design-only oracle before invocation; the lease helper now waits for its own token, and the final proof uses a fresh label. That failed phase is preserved without treating the non-invoked oracle as a test pass. A later remote lease read encountered an in-progress JSON write and terminated only its own test. Remote lease writes now fsync and atomically replace the file; the interrupted mutation remains a failed phase, and fresh labels cover the final complete verification.

Measured bitmap boundary and repair

Evidence: scratch/runs/20261102-gl-bitmap-boundary-contract/ (gitignored). The current ordered Oni CLI dependency and Brave causal-observer dependencies completed before source ownership and execution. Actual Chrome 151 / ANGLE Intel UHD 620 / Mesa 23.2.1 WebGL1 and WebGL2 ran the same 3×2 literal controls before and after the repair, with non-pixel row padding set to 0x6d.

On unchanged source, 32-bit BGRA words 0x3c283848, 0xff283848, 0x0096aa55 imported as RGBA [40,56,72,255], [40,56,72,255], [150,170,85,255]: exposed alpha 60 and 0 were lost. A read-only access followed by publication, or a point with all color writes disabled, changed canonical alpha bytes to 255. An alpha-only point produced attachment alpha 128 which publication also replaced with 255. RGB and row padding were unchanged. The unchanged-old faithful oracle exited 1 for these literal mismatches.

RGB555 words 0xa8c7 and 0x28c7 both imported as [82,49,58,255]. Their high bit is unexposed padding, not alpha: resolved native queries report RGB widths 5/5/5 and alpha width 0 (32-bit reports 8/8/8/8). Old publication cleared the high bit even after read-only or zero-mask use. An alpha-only draw also exposed private RGBA8 attachment alpha 128 despite that zero-alpha-bit format. These are separate channel and storage contracts. No-access publication left canonical bytes intact on old and repaired source.

The repair opts GL bitmap transfers into preserving 32-bit alpha and bits outside the 16-bit RGB masks; ordinary GDI conversions retain their existing opaque defaults. WebGL publication reads the drawable attachment directly, flips its rows, and avoids the opaque presentation canvas. Bitmap transport carries bpp to the renderer; RGB555 suppresses writes to private attachment alpha. RGB565 has no unused high bit, verified by its red-bit control.

Literal unit and actual WebGL controls cover unchanged pixels, all-zero masks, RGB-only and alpha-only writes, alpha-zero RGB, row orientation/padding, aligned and unaligned storage, and fresh GDI writes after publication. Five mutation controls independently fail for discarded import alpha, store alpha, unused bits, opaque presentation readback, and exposed private RGB555 alpha. Resolved native direct and actual queued software tests cover 16/32-bit DIBs in both row orientations, stack/register preservation, masks, literal stores and queries; the queued run recorded ten GL submissions with zero fallbacks. Worker bitmap transport has a focused protocol test; this is not an actual queued WebGL test. The native resolver test explicitly requires glLogicOp to remain unavailable. Focused GDI surface, bitmap-target, frame-state, shared-render-worker and software-producer regressions passed. Procedure availability passed with all three intended lookup mutations rejected; the scratch preload only disables fonts and records the unchanged compiled bytes. The locked full build passed.

The first guarded execution phase hit the disk floor at 1,996,500,992 bytes and terminated its own native test (exit -15); it is not accepted as a passing test. Evidence retains that breach and its cause is unknown. Execution resumed only after free space recovered above the floor and a second watcher was started. That watcher was absent at final audit without a terminal minimum record; continuous floor coverage of the recovered phase cannot be certified. The initial full build and availability test stopped for omitted font assets, not semantic failures; their original errors remain in the run. No game fixture, download, deployment or original-game replay was used.

At the boundary-repair handoff: implement the sixteen logic operations against stored component widths, with explicit 5-bit source/destination quantization for RGB555 and independent RGBA8 channels for 32-bit surfaces. At that handoff WebGL still used an RGBA8 backing attachment; this boundary repair does not establish low-bit logic quantization, dithering, overlapping primitive ordering, independent front/back destinations, or the full state/error/list contracts. Use non-feedback destination snapshots and literal destination-dependent oracles before changing availability. This run does not establish a WebGL implementation impossibility.

Earlier blocked semantics run

That earlier run made no runtime, dispatch, transport, backend or test changes. Its evidence remains under scratch/runs/20261102-gl-logic-op-semantics/.

The ordered predecessor gates passed: Oni93804 terminal/result/final/exact AUTHOR RELEASE and 70 final file hashes, then Brave99598 and 364 final file hashes, with named and scoped helpers absent. Main was refreshed to d64373af1ea6a97df7845e250200bba240bd3b39. Foreign and laptop WIP were preserved.

Concrete execution blocker

The guarded direct source stream exited 1 when /home free space fell to 1,961,848,832 bytes; its exception cleanup observed 1,961,193,472 bytes. Both violate the task's 2,000,000,000-byte floor. The initial source guard also reserved 4 MiB. Available space later recovered, but this run did not retry the transfer or claim the floor held. The cause of that transient allocation is unproven. stream-source.log preserves both exceptions; execution-abort.json records the abort and stream-inputs.json.gz pins the intended files by size/SHA-256. No local archive was created.

Own finite-TTL boat bx_3a2fwcqj had original expiry 2026-11-02T23:58:05.494Z (TTL 7200 seconds), recorded in box.json and on the messageboard before remote use. The stream aborted before extraction. Remote abort inventory and its verified retrieval, cleanup and explicit delegated shutdown are recorded in the run. STOPPED was verified at 2026-11-02T22:02:45.763Z, with remote helpers absent before shutdown and local poll/transfer helpers absent afterward. The partial source archive is not an accepted build input. The initial inventory command failed on the ps header; the corrected headerless inventory and original error are retained separately.

Not run: unchanged-old API failure control; native resolved API tests; any of the sixteen destination-dependent pixel cases; actual queued software; WebGL1/2; semantic mutation controls; focused regressions; full build. No module, rendered pixels, backend-format proof or original-game observation was produced.

Contract and implementation work still required

Primary references read:

Color logic must operate independently on each selected destination's stored component bits after source processing and quantization; write masks follow. It overrides blending without changing requested BLEND state. INDEX_LOGIC_OP is distinct and has no effect in RGBA contexts. Initial mode is COPY with both enables disabled. Invalid operations and Begin/End misuse must preserve state. Clear and accumulation RETURN bypass logic. Context, display-list, attribute and queued-state ownership must remain correct.

Read-only inspection found native 16/32-bit stores in src/09ab-handlers-d3dim-core.wat, native state/snapshot/bitmap paths under src/09a8*, and production transport/frontends/backends in lib/. The shared native final store must retain correct 565/555 packing and padding. WebGL needs separate destination copies, with ordering for overlapping primitives and independent front/back results; sampling the drawing attachment is invalid.

A specific unresolved format issue needs an actual driver test: WebGLBackend.createDrawable currently creates RGBA/UNSIGNED_BYTE textures, while a native bitmap context can target a 16-bit X1R5G5B5 DIB and report five RGB bits. An eight-bit logic test cannot establish that bitmap contract. Whether an actual low-bit attachment can faithfully support it on each WebGL route is unverified, not a demonstrated backend impossibility.

Unexecuted helper, shader and resolved-native-harness drafts remain in the run; they are preparation only. They must be reviewed and completed before use. The public unsupported gate must stay intact until the full requested semantics pass, including literal destination-dependent results, format/mask/error/state and ordering controls, actual software queue with zero fallback, actual WebGL1/2 routes, meaningful mutation failures and a locked full build.

Recovery

Restore stable disk headroom above the required floor before fresh execution; do not reuse the stopped boat or the incomplete transfer. In a fresh finite-TTL boat, record its ID/expiry before use and stream exact hashed source without a local archive. The retained stream-source.py and remote.py are recovery recipes; adjust the newly recorded box, and inspect their input scope and disk guards first. Complete the old-source control before implementation, then the full semantic and regression work above. Retrieve/SHA-verify evidence before any explicitly delegated own-boat shutdown. This note supplies no basis for removing glLogicOp from the unsupported-API test.