feat(sidebar): hover details card for workspaces

[Zeus-Deus]

Why

Every sidebar surface is lossy. An inbox card truncates its title and never renders git_behind or git_changed_files at all — both already ride on WorkspaceSnapshot and were displayed nowhere. A settled row shows only a title. A collapsed rail avatar shows nothing but a status dot, behind a title-only tooltip.

Resting the pointer on any of the three now opens a shared read-only HoverCard to the right (350ms open, 290px, side="right" align="start").

What it shows

Header — repo eyebrow, provider marks, agent state with elapsed, the full untruncated title, and the linked issue's number + title.

Rows, all conditional except Location — Branch, Uncommitted +A −D, Changed files, Ahead, Behind (or a single "Working tree · clean" row when there is nothing to report), Pull request, Issue, Port(s) capped at 3 with a +N overflow, Location (This device / host name / host · in place), Notifications muted.

Footer — the real path on disk (worktree_path ?? remote_cwd ?? cwd), $HOME collapsed to ~, wrapped rather than truncated.

No new Tauri command. Every field is already on WorkspaceSnapshot or in appState.detected_ports. Styling reuses the DetailRow label/value shape from the context bar's popover rather than inventing a second vocabulary.

Implementation notes

• The body is a separate component so Radix's unmount-when-closed keeps a sidebar of N workspaces from subscribing N times to the ports/hosts stores at rest.
• Ports are selected as a raw slice and narrowed in a useMemo. Filtering inside the zustand selector returns a fresh array per call, which breaks useSyncExternalStore snapshot equality and spins React into an infinite re-render loop — this crashed the card tree during development and was caught only by running the real app.
• Remote-ness keys off host_id alone. The hosts list loads asynchronously, so an unresolved name falls back to "Another device" rather than letting a failed lookup claim the workspace is local.
• Mounted inside WorkspaceInboxMenu on a different node than ContextMenuTrigger, so the two asChild triggers never compose onto one element.
• The label holds its width and the value truncates. The reverse rendered long worktree branches as Bran… next to the thing you came to see.
• The card is pointer-events-none — a detail surface, not an interactive one. Reviewers may want to overrule this: it means you cannot mouse in to select the branch name or path.

Removes the rail's title-only tooltip and the card's redundant native title attribute, which would have raced the hover card on the same element.

Verification

• npm run check clean.
• npm run test — 210 files, 3069 tests pass, including 19 new ones covering the git rows, clean-tree fallback, non-git workspaces, PR/issue tones, port capping, the async-host fallback, mute, and path shortening.
• Visually confirmed in the dev build across all three surfaces (inbox card, settled row, collapsed rail), with instrumented pointerenter proving real mouse input opens it.
• Zero raw Tailwind palette utilities — status/accent tokens only, per the design system.
• npm run verify does not complete in this worktree: cargo check fails on a missing src-tauri/binaries/agent-browser sidecar. Pre-existing environment gap, unrelated — this branch touches no Rust.

Docs updated in docs/features/sidebar.md (inbox section, rail section, What Works, Constraints, Touch Points).

[Zeus-Deus]

Review — Opus 5 (high effort)

Solid, and the two hardest details are handled correctly:

• The ports-selector trap is genuinely avoided. Selecting the raw slice and narrowing in useMemo is the right shape; a filtering selector really would break useSyncExternalStore snapshot equality and spin. Good that this is written down in the code, not just the PR body.
• The two asChild triggers are on distinct nodes in all three call sites, which is the failure mode I'd expect from this composition.
• Remote-ness keying off host_id alone with an "Another device" fallback is the right call given the async hosts load.

pointer-events-none contradicts the footer's own rationale

{/* Real path on disk, last: the least-scannable line, and the one users
    most often want to copy or confirm. Wraps rather than truncates. */}

…inside a card that is pointer-events-none, so it cannot be selected or copied. You already flagged this for reviewers, so: I'd overrule it. The path and the branch name are exactly the two values a user reaches for a hover card to get, not just to read. Radix HoverCard handles mouse-into-content fine (that's what closeDelay is for), and dropping the class costs nothing except that the card intercepts clicks in its own footprint — which is over the main pane, not over other sidebar rows. If it stays non-interactive, the footer comment should stop claiming users want to copy it.

Settled rows always report "Idle"

SettledRow passes status={null} unconditionally, so the header renders Idle even for a settled workspace whose agent is actively working — and #206 makes "settled but working" a real, reachable state (its auto-un-settle fires on working). The row itself shows no status either, so this isn't a regression, but the hover card is the surface whose whole purpose is to un-lose what the row dropped. getWorkspaceStatus(workspace.surfaces, paneStatuses) is already available.

Two smaller notes

• elapsed freezes at open. formatElapsed(Date.now() - stateSince) is computed once per mount, deliberately. Fine for a 350ms-to-open card you glance at — but it's the one value on the card that visibly goes wrong if you leave the pointer resting, and "Working 4m" that stays 4m for two minutes reads as a bug.
• The bare <div> wrapper in SettledRow now sits between the settled list's flex container and the row, so it becomes the flex item. Almost certainly fine, but it's an unlabelled layout node — worth a contents-style comment or className="contents" if the row was relying on being a direct child.

[Zeus-Deus]

Review by GPT 5.6 Sol xhigh found a path-boundary bug.

shortenPath at workspace-hover-card.tsx:48-50 uses raw path.startsWith(homeDir). That treats sibling prefixes as descendants: with home /home/u, a workspace at /home/u2/project is rendered as ~2/project. The same issue occurs with trailing separators and platform-specific roots.

Please compare path components (or require the next character to be a separator, while handling exact-home as a special case) and add the sibling-prefix regression. The focused hover-card suite passed 19/19 tests; it currently covers an actual child and a completely unrelated path, but not this boundary.

Coordination note: this head has pairwise textual conflicts with #206 and #213.

[Zeus-Deus]

Verified: all five review issues are genuinely fixed at the head (pointer-events restored on the card, settled rows pass real status, elapsed ticks via the shared coarse clock, shortenPath sibling-prefix/exact-home/trailing-separator cases covered). tsc clean and all touched suites pass (23/23 hover-card, 30/30 sidebar-inbox, 6/6 rail).

Blocked on a rebase now that #206 merged — conflicts in sidebar-inbox.tsx, sidebar-inbox.test.tsx, sidebar-inbox-card.tsx, sidebar-rail-workspaces.tsx, docs/features/sidebar.md. Should be mostly textual; one semantic point: main's settling work adds new tiers/states for settled rows, and this PR's settled-row status prop should reflect those (the hover card anticipated exactly this, so integration should be natural).