Switching OpenGL → raytracing → OpenGL could leave the viewer showing only the
background gradient, with mid-session "[gl1] WebGL context changed" thrash and
INVALID_OPERATION storms on BOTH WebGL contexts. Traced mechanism: the shim's
FFP draw routing keyed on one process-global client-array flag; an interrupted
fixed-function window (MODEL_3D::BeginDrawMulti loops, or the switch-back
reload re-recording display lists with the GL context lock released across
JSPI suspensions) left it set, after which the raytracer blit's and the 2D
GAL's glDrawArrays were routed through the FFP pipeline — and one misrouted
draw permanently repointed the VICTIM's own VAO attributes at shim buffers
(the blit's attribute 0 collides with ATTR_POSITION), so both stayed
broken/blank even after the flag cleared.
Shim fixes (wasm/gl1):
- Owner-context routing gate: __wrap_glDrawArrays/Elements route into the FFP
pipeline only under the shim's owner context (adopted at the first FFP
client-state mutation or programSync in a context); foreign-context draws
always pass through — the 2D GAL can never be misrouted and the context
guard can never thrash.
- VAO isolation (ScopedDefaultVAO): draw executors do their attribute setup on
VAO 0 and restore the caller's binding — a misrouted draw can no longer
corrupt the caller.
- contextSync() resets the whole client-array mirror on a context change
(enables/pointers/VBO names all described the dead context).
kicad pointer bump (d6e3dc1a87a): blit preamble disables the four client
arrays (same-context firewall) + DoRePaint hidden-parent early return now
clears m_is_currently_painting like its six siblings (a standalone
sufficient cause of a permanently blank viewer).
TDD (each observed red before its fix, green after; harness = authoritative):
- T1 VAO corruption, T2 foreign-context routing + guard thrash, T3 stale
client-state surviving context recreation — tests/e2e/3d-webgl.spec.ts over
new harness choreography (appQuad/ffpMakeStale/createSecondContext/
quadDrawFresh). Parity stays 47/47 with zero drift.
- tests/kicad/3d-viewer-engine-toggle.spec.ts (new, CI-skipped like the
deadlock spec): real round-trip happy-path gate — board re-renders, zero
[gl1] lines, zero INVALID_OPERATION. (The raytraced image itself never
displays on the wasm build — the pre-existing inert-toggle KNOWN ISSUE in
3d-viewer-deadlock.spec.ts, out of scope here; the engine switch and the
poisoning reload path run regardless.)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Closing the viewer destroys its wxGLCanvas's WebGL context; reopening mints a
new one. The gl1 shim cached GL names (FFP program, stream/scratch VBOs) in
never-reset statics behind `if (!handle)` guards — in the new context every
draw died with INVALID_OPERATION and the viewer showed only the clear color
("Reload time 0.031 s" is benign: warm model caches make the rebuild fast).
contextSync() (gl1_state.cpp) now detects the context change in programSync()
— the one choke point every shim draw crosses, and a path the 2D GAL never
reaches (a first attempt checking in the glBindTexture wrap saw the GAL's
context and thrash-rebuilt the program 23x per run) — and drops the cached
names for lazy rebuild in the new context. Context identity is a monotonic id
stamped on Emscripten's per-context record: the numeric
EMSCRIPTEN_WEBGL_CONTEXT_HANDLE is recycled, so a destroy-then-create can
return the same number and a handle comparison detects nothing.
The lost-position half is a wxwidgets wasm fix (pointer bump: GetFromWindow
reports display 0; saved geometry used to carry display=(unsigned)-1, which
LoadWindowState treats as "display not found" and re-centres the frame).
TDD (red observed before each fix, green after):
- tests/kicad/3d-viewer-reopen.spec.ts (new, own worker like the deadlock
spec): load board, open viewer, render-gate, drag by the titlebar, close
via the x, reopen; asserts the board re-renders (was: 1 distinct colour for
90 s) and the window position is restored (was: re-centred to 0,0 after
closing at 40,90). Green run logs exactly one [gl1] context-change line.
- 3d-regression harness: recreateContext() destroys the context AND swaps in
a fresh canvas element (a browser canvas keeps its context for life, so
same-element recreation hands back the live old context and hides the bug);
the new 3d-webgl spec test renders redraw-mini-board-navigator before and
after recreation and requires pixel-identical output. Parity: 47/47, zero
drift.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
tests/3d-regression mirrors the gal-regression pattern at renderer scale:
shared C++ scenarios call real KiCad 3D-viewer code (opengl_utils, display
lists + DrawCulled stencil subtraction, MODEL_3D VBOs, private generators via
a rob-template accessor, and full reload()+Redraw() composites over a
synthetic BOARD_ADAPTER). A native macOS harness renders them on real OpenGL
into 47 committed goldens (bit-deterministic, FBO capture); the wasm harness
compiles the same TUs against wasm/stubs/gl_ffp_stub.c no-ops so every
scenario renders blank — the TDD red state (parity meter: 47/47 changed).
Comparisons use the CI pixelmatch engine via the new generic compare-dirs.ts
(floors.json levels; manifest.json cmp-guards registry drift).
Documents an upstream bug: appendPostMachiningGeometry's countersink path
adds middle quads without normals, silently erasing the walls of any
display list it is batched into (3d-post-machining.png keeps the lists
separate to record it).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
