Upstream resynchronization assessment

This is a cost assessment, not an import plan or an update receipt. GPUI Box remains the development authority, no upstream tree was merged, and the frozen vendor refs were not advanced.

Assessed coordinates

The assessment was generated on 2026-09-04 with:

scripts/sync-zed/sync-zed assess \
  --revision 801c087af22dd189dc1aa49e2f370b4f04190b19

Coordinate	Revision
Frozen filtered base (`vendor/zed-gpui`)	`c036e5bcb472b7c557c231a66d69e646285d1942`
GPUI Box `ours`	`c801eecb34e0448f8a23b7a3d8394a76b48b322c`
Zed source snapshot	`801c087af22dd189dc1aa49e2f370b4f04190b19`

The Zed source is 483 commits ahead of the official frozen baseline a6a23c7b80a5cefa0487b7856335be89ace7e483. The frozen path filter produced 327 files. All 18 configured package manifests were present and all 18 package names were rewritten to their GPUI Box identities before merging. The JSON report's SHA-256 is 30b85f08c93bb864581bed8465e47297764d822393045c5238fd8f1fd5b3a1fa.

The command fetched and merged only inside a temporary repository. A hash of the current repository's complete show-ref output was identical before and after the run.

Measured merge surface

Measure	Files	Additions	Deletions	Changed/payload lines
Filtered upstream change against the frozen base	152	17,548	11,358	28,906
Automatically merged, fully conflict-free files against `ours`	93	9,029	2,373	11,402
Conflicts	57	—	—	8,287

There are 196 conflict hunks. Five unmarked modify/delete or binary-style conflicts count as one hunk and zero payload lines each. For text conflicts, “payload lines” counts both sides between Git's conflict markers and excludes the marker lines. The 11,402 automatic-merge lines count only staged files with no conflict anywhere; automatically merged portions of the 57 conflicted files are deliberately excluded. These measures therefore describe review surface and must not be added together as a patch-size total.

Conflicts by subsystem

Classification uses this precedence: platform_view, accessibility, platform renderer, scene, window, text, elements, then other/Kit-unrelated. The frozen filter contains framework and support crates, not gpui-kit, so the last category means that a conflict has no direct Kit source ownership.

Subsystem	Files	Hunks	Payload lines
Window	5	48	1,383
Scene	1	20	1,558
macOS renderer	3	3	0
Windows renderer	1	4	559
WGPU renderer	5	33	2,242
Web renderer	0	0	0
Linux renderer	0	0	0
Other renderer	1	1	19
Elements	3	20	919
Text	1	1	7
PlatformView	0	0	0
Accessibility	1	1	79
Other / Kit-unrelated	36	65	1,521

Ten largest conflict files

File	Subsystem	Hunks	Payload lines
`crates/gpui_wgpu/src/wgpu_renderer.rs`	WGPU renderer	25	1,594
`crates/gpui/src/scene.rs`	Scene	20	1,558
`crates/gpui/src/profiler.rs`	Other / Kit-unrelated	6	988
`crates/gpui/src/window.rs`	Window	36	891
`crates/gpui_windows/src/directx_renderer.rs`	Windows renderer	4	559
`crates/gpui/src/elements/div.rs`	Elements	4	535
`crates/gpui_macos/src/window.rs`	Window	8	444
`crates/gpui_wgpu/src/gpui_wgpu.rs`	WGPU renderer	1	422
`crates/gpui/src/elements/animation.rs`	Elements	15	252
`crates/gpui_wgpu/src/shaders_storage.wgsl`	WGPU renderer	3	158

Already adapted versus absent

This inventory is based on local provenance, local commit history, and the present code, not patch-id ancestry: the existing ports intentionally use GPUI Box types, names, tests, and behavior boundaries.

Zed source revision	Local status	Evidence and boundary
`8b1497dbd22fb06f5838a7c0b84a1e54fafa71bc`	Adapted	Local `1a37ce5d2017e48acae854563275d762e22db36d` added the framework spring solver, animation lifecycle, and Kit delegation. Zed's product examples were not imported.
`956a49e4ca8aa4b7c2c293e1414c91f009824ae3`	Adapted	Local `4bf127119f4ef7a65a4cb15d99d7fc2f250af999` added raw touch and portable recognition at the framework boundary.
`76b1096cbd83b5b5138793e5f552218abc8fdcbb`	Adapted	The same local gesture port carries platform-selectable exponential and Android spline scroll physics.
`0855410ccd2040efbbf14d71409166b6c472e0bd`	Adapted	The same local gesture port carries long-press recognition and claiming.
`b3326e13c142fc8f313aca67a93dd6855a1e7e32`	Adapted	The same local gesture port carries phased `TouchDragEvent` dispatch.
`5e28272c1407ced4bae4a90deaea25352a1fbc96`	Adapted	The same local gesture port locks a pan to its initial dominant axis.
`a21007b7a948e46afbe719150f5e9968bfcd1078`	Partially adapted	Local `2431dc84dc2ae25ae02b809057fd2a196576d09c` retained GPUI Box's profiler authority and added the selected draw/submission hooks; it did not import Zed's profiler UI and product wiring.
`9e236090b9a31338caf233d440f724922b58d7e1`	Partially adapted	Per-window frame timing informed GPUI Box's bounded `FrameTimingMonitor`, but upstream histogram/product dependencies were not imported.
`1861e58f984c76afc06032e753557994ffc8fe44`	Deliberately absent	Foreground journals, hang incidents, telemetry, and their product integration were explicitly excluded.
`55007f518bc1d49e6b3291c5eaa1aabf649b36fd`	Semantically narrowed	GPUI Box reports dirty-to-submission, because return from `PlatformWindow::draw` does not prove compositor/display presentation. It does not claim or import dirty-to-present behavior.

The four additional fixes called out for this review have mixed status:

Zed revision	Local status	Evidence
`eb354c8d504071bdb79110a7a5c9d374c2864113`	Absent	The current Wayland path still has `renderer_presented` and `completed_frame`; it does not have Zed's `schedule_frame`/`FrameLoop` demand-driven state machine.
`ae99a867d7a24682435bd1821c66b4e172a10768`	Present in the frozen bootstrap	Both the frozen vendor tree and current X11 path coalesce `Expose` windows, ignore unmapped windows, and immediately request `require_presentation`. This was not a post-freeze manual port.
`511ac170363776319b38cc0e9c047a06aa2e7541`	Absent	Inactive-window throttling remains hard-coded to 33,333 µs; `WindowOptions` has no `inactive_frame_interval`.
`a85cf449ec6e72274ee6f9e8462f702299b04cce`	Adapted	Despite being a touch-prediction fix rather than a render-loop fix, its `last_movement`/opposing-vector guard, non-reversing momentum, and regression test are present in local gesture port `4bf127119f4ef7a65a4cb15d99d7fc2f250af999`. P18 records the reviewed target revision but does not list this follow-up separately in `source_revisions`; this assessment does not mutate that frozen source receipt.

This is a scoped inventory of the cited ports and fixes, not a claim that the remaining commits in the 483-commit range were reviewed individually.

Cost comparison and decision

Continuing selective ports keeps each framework change product-neutral, reviewable, attributable to exact source commits, and testable on the affected platforms. It also avoids automatically adopting product assumptions and upstream workspace dependencies. Its recurring costs are source archaeology, translation into diverged APIs, separate provenance, and the risk of missing a companion fix. The mixed render-loop inventory above demonstrates that risk.

A discrete whole-tree import would preserve more upstream relationships in one operation, but it is not mechanical: 57 conflicted files contain 196 hunks and 8,287 payload lines, concentrated in the fork's most changed renderer, scene, window, profiler, and element authorities. The 93 cleanly merged files still add 11,402 lines of semantic review surface. All affected Linux, macOS, Windows, Web, accessibility, renderer, generated-API, dependency, and headless contracts would need validation after conflict resolution; pixel-changing renderer decisions would additionally require native baseline review.

The measured near-term choice is therefore to continue deliberate manual ports. A discrete overall import should be treated as a separately staffed migration and should begin only after explicit policy and provenance sign-off, not by applying this assessment's temporary merge result.