[Juliusolsson05]

Problem

The newly enabled GPU path exposes corruption in pinned xterm WebGL 0.19.0 under colored CLI output and scrollback. Atlas merges during rendering leave stale cell coordinates; per-page texture-version collisions can also keep the wrong GPU texture bound. Refresh alone is insufficient.

Change

• Coalesce atlas layout events into one deferred texture-binding invalidation plus viewport refresh through xterm’s frame scheduler.
• Preserve GPU acceleration, shared glyph cache, scrollback, terminal dimensions, palette, and PTY/input ordering. No periodic or unconditional per-scroll repaint.
• Fence pending repairs and dispose all subscriptions on context loss/unmount.
• Isolate one shape-checked, version-pinned private bridge in the loader: setAtlas with the existing atlas invalidates GPU bindings without clearing glyphs. Fail back to DOM if this bridge is incompatible or throws. Remove it when a stable addon includes upstream xterm.js #5883.

Refs #789
Upstream: xtermjs/xterm.js#5883 and xtermjs/xterm.js#6038

Verification

• Type-check passed.
• 25 focused tests passed: version canary, coalescing, invalidation ordering, teardown/context-loss fences, partial registration failure, DOM fallback, and mounted shell/agent/inline host behavior.
• Test-contract check and git diff --check passed.
• Repeatable real Electron GPU script: node scripts/smoke-terminal-renderer.mjs; use --control for the unwrapped addon.
• Control: 33 output-frame mismatches, one scroll mismatch, 1,264,868 differing pixel channels against a fresh independent atlas. Production loader: zero frame mismatches and zero reference differences; GPU stays active.
• Synthetic screenshots inspected. No live Agent Code restart, provider calls, session termination, or user profile access.

Limits

The exact user screenshot still requires post-deployment confirmation, so this references rather than closes #789. The fix repairs stale frames on the next scheduled frame; it is not an upstream synchronous merge-loop backport. GPU smoke coverage is opt-in because ordinary headless CI cannot guarantee a usable graphics context. The broader React/performance refactor remains isolated in its separate worktree. Not merged; awaiting review and CI.