Iced: upstream-blocked drop/DnD and Wayland clipboard gaps (roadmap 3g)

[charliek]

Tracking issue for the drop/DnD gaps that are upstream iced/winit limitations (plan 015 W4). Discipline pinned in plan 013/015: track and document — no in-process workarounds, no upstream engagement.

The four gaps:

1. No drop coordinates. winit surfaces file drops only as window::Event::FileDropped(path) with no cursor position (crates/roost-iced/src/main.rs maps it 1:1; a unit test pins the coordinate-free shape). Exact drop hit-testing is impossible in-process; roost targets the active terminal when no palette/editor owns input (native_file_drop_origin, crates/roost-iced/src/app/interactions.rs) — that stays.
2. No text/URI-list drop events. winit has no raw-text or URI-list drop event at all, so text drops cannot be received. (Adjacent, same root: clipboard text/uri-list is also unreachable through the string-only/image-only clipboard stacks, so a file-manager image copy that sets only text/uri-list pastes under GTK but not under iced — plan 015 C6 documented divergence.)
3. No Wayland DnD. The iced/winit stack delivers no file-drop events under Wayland — native Wayland DnD is absent entirely.
4. Wayland clipboard seat serial. Clipboard writes need a seat serial the stack does not thread; the iced clipboard e2e tier is skipped under Wayland (Makefile e2e-iced/e2e-iced-ci gate on WAYLAND_DISPLAY; .github/workflows/ci.yml documents the weston headless seat limitation).

Related but distinct: #101 tracks the GTK-side Wayland cross-process clipboard issue; #282 (closed) was drop-path content filtering.

Exit condition: an iced/winit release that delivers any of these becomes its own adoption slice; until then this issue is documentation.

[charliek]

Re-investigated 2026-08-10 against current upstream state (plan-024 close-out). Three of the four gaps have materially changed, and one was mis-framed from the start. Roost pins: iced =0.14.0 (unpatched crates.io winit 0.30.13), window_clipboard 0.5.1, smithay-clipboard 0.7.3, arboard 3.6.1.

The big upstream event

winit merged PR #4571 "New drag and drop API" (2026-07-16, master): a full DataTransfer API with DragEntered/DragPosition carrying cursor position, MIME-typed fetch (TypeHint::Plaintext | UriList | Html | …), and a real Wayland backend (winit-wayland/src/dnd.rs, SCTK DataDeviceManager, receive and send). The 5-year-old winit #1881 (Wayland DnD) closed as completed 2026-08-03.

None of this is released. Latest stable is 0.30.13; the 0.31 betas predate the merge by 8 months; the 0.31.0 milestone has 27 open issues and no date. And iced 0.14 does not patch winit, while iced master pins a fork that is still 0.30.8-based — so no iced release moves winit today either.

Per-gap corrections

Gap 1 (drop coordinates): blocker moved from winit to iced. winit master has the position. But iced's in-flight winit-0.31 upgrade (draft PR #3415, opened 2026-08-10) maps the new events back onto the existing FileDropped(PathBuf) API — it discards position entirely. So even post-upgrade iced gives us no coordinates without a further API-breaking iced change, which nobody has proposed (iced #1642 is the standing complaint). Still a product gap.

Gap 2 (text/URI-list): split it — half is not upstream-blocked.

• Drop half: winit master supports TypeHint::Plaintext/UriList; iced #3415 fetches UriList only and drops plaintext on the floor. Product gap, blocked on iced.
• Clipboard half: smithay-clipboard 0.7.3 hard-codes 3 text MIMEs and window_clipboard is String-only — but roost can read text/uri-list directly via wl-clipboard-rs (the same zwlr_data_control protocol arboard already uses for the focus-free image read) or x11-clipboard, exactly as the GTK UI does through GdkClipboard (crates/roost-linux/src/clipboard.rs:79). That is a roost scope decision, not an upstream blocker. This issue's original framing conflated the two halves.

Gap 3 (Wayland DnD): upstream-solved, unreleased — and the one that lands "for free". iced #3415 needs no public-API change for this half: with a winit-0.31 backend, FileDropped simply starts firing on Wayland and roost's existing 1:1 mapping works unchanged. Roost-side change when it ships: delete the "Wayland: not implemented" caveats and re-enable coverage. Watch list: winit 0.31.0 milestone → iced #3415.

Gap 4 (clipboard seat serial): the framing here was wrong — this is a TEST gap, not a product gap. smithay-clipboard 0.7.3 does thread the serial (seat.latest_serial → set_selection, gated on has_focus — read at tag v0.7.3, src/state.rs). The refusal happens because headless weston has no input seat, so no serial exists — a Wayland protocol rule (an unfocused client must not hijack the clipboard), identical under GTK (#101 family). The product path is already proven on a real seat by tools/input/linux/iced_wayland_clipboard_check.py (cage + uinput); what's missing is that harness in CI. Filed separately since it has no upstream dependency at all. No upstream release will ever change this one. (One narrow protocol-mandated edge survives: an OSC 52 write while the window is unfocused is silently dropped — same under GTK, documented behavior.)

Evaluated and rejected: the COSMIC fork stack

System76 has all of this working today — but only inside pop-os/iced + pop-os/winit + pop-os/window_clipboard (with git-only dnd/mime crates; the crates.io dnd is an unrelated RPG crate). Adopting it means moving off upstream iced entirely onto a vendored fork chain, against this repo's released-and-exact-pinned discipline. Named here so it isn't re-investigated.

Updated exit condition

Watch winit 0.31.0 stable and iced PR #3415. When both land: gap 3 is a version bump; gap 1 and the drop half of gap 2 additionally need an iced API change no one has proposed. Gap 4 exits via a roost CI lane, independent of all of the above.