A card that hasn’t arrived is not a card that left

Why a Headway deep link opened and instantly snapped back to the board — and the one field that tells the two cases apart.

headway:notedeck/super-hello-emotion · commit e757177ea8d8 · epic headway:notedeck/obey-margin-scale (2 of 5)

The loop that produced the bug

Headway doesn’t own its navigation. Each frame the chrome hands it a route token through render_nav; Headway seeds state.selected from it, renders, and then diffs where the UI ended up against where the route said it started. That diff — reconcile_nav — is what turns a click into a global-history push, or a dismissal into a Back.

So anything that quietly clears selected mid-frame is indistinguishable, to the diff, from the user pressing ✕.

One frame of Headway

What the user saw

jb55 reported that clicking an inline headway: reference from another app needed two back presses, the first landing on the board root. Commit 1 of this epic explains the doubled push. This card is the other half: even once the push is single, the opened card wouldn’t stay open. A cross-app deep link pushes the routed entry before Headway has necessarily folded that card in from nostrdb — so find_card misses on the very first frame, the selection is cleared, and the entry the chrome just pushed is popped by the app itself.

The discriminator: `detail_for`

Both failure and success look identical from selected alone: a NoteId that find_card can’t resolve. The field that already distinguishes them is detail_for — the card the detail pane’s edit buffers were last seeded from. It’s set the first time the pane actually renders a card, and cleared when it closes. So:

state	meaning	right move
detail_for == selected	we rendered this card here; it has since left the board (deleted, archived, moved away)	clear → emit `Back`
detail_for == None	the pane never showed it: the card hasn’t arrived yet (fresh deep link, remote card in flight)	hold → emit nothing

-    if state.selected.take().is_some() {
-        state.detail_for = None;
-    }
+    if state.detail_for.is_some() && state.detail_for == state.selected {
+        state.selected = None;
+        state.detail_for = None;
+    }

Held costs nothing visible: the board grid draws underneath, and the detail opens on the frame the card’s events land. Nothing on the grid render path reads selected — the only write is the click handler, which overwrites it anyway.

Deliberate asymmetry

The graph_epic branch directly above still drops unconditionally, and now says why in a comment: a dependency graph is only ever entered from an epic already on the board. No deep link mints a Graph route, so there is no hasn’t-arrived-yet case to wait for.

Measured, both directions

Two tests in crates/notedeck_headway/tests/snapshot_tests.rs, both behavioural (no lavapipe). The regression test sets nav_token to a Card route for an id on no board, pumps five frames, and counts NavRequest::Back:

Over-suppression guard

Holding an unresolved selection must not swallow the real dismissal. The second test opens a card’s detail for real, clicks Delete card, and asserts exactly one Back. It works because resolve_detail_outcome clears selected itself in the same frame, so the Card→Board diff still fires without depending on find_card missing. That path had no test before this commit.

Not verified here

No real cross-app deep link was exercised. That’s commit 5 (headway:notedeck/duck-echo-chronic). This commit removes a blocker for it; the end-to-end “one back press” claim is only provable there.
The held-then-resolved transition isn’t directly tested. The regression test proves no Back is emitted while the card is absent; it does not ingest the card mid-test and watch the detail open. Worth adding when commit 5 lands a real deep-link test.
Stale-token behaviour under the real chrome. In this chrome-less harness the token stays Card(deleted) after a delete, which would re-seed the selection on the next frame. Under the real chrome the Back is drained and the token becomes Board first. The delete test therefore only asserts the single frame after the click.
Snapshot images not regenerated — none of the 22 lavapipe snapshots are affected (no visual change on a board with all its cards present).

Gates run

./scripts/ci-local — changelog-check, lint (fmt + clippy), linux-test, android, snapshot-test: all passed.
cargo test -p notedeck_headway --test snapshot_tests — 10 passed, 0 failed, 22 ignored (was 8 passed).