Branch picker rows, unread panes, and two Markdown wrapping fixes

[matej21]

Ten small, independent fixes and features that accumulated on one branch. Nothing here shares code, so the commits are reviewable one at a time; they are grouped below by the area they touch.

Branch picker

The picker listed bare branch names in git's ref order. Choosing a branch meant remembering which one is current work, whether it is pushed, and whether a worktree already holds it — the last one only surfacing as a failed checkout.

• The current branch leads the LOCAL section and becomes the default keyboard selection. The filter/order pass is extracted out of recompute_branch_filtered into a pure branch_nav_items so the ordering is unit-testable.
• Every row carries its state, right-aligned: the worktree holding it, local / gone / ↑N ↓M against the upstream, and the tip age. Sections order by recency. All of it comes from one git for-each-ref pass — git answers ahead/behind for every ref off its own commit-graph (~10ms for ~70 refs here), where a per-branch gix rev-walk does not come close. It rides the existing BranchList wire type as an additive field, so a remote host without it degrades to plain names.
• A row context menu with the three things that need no checkout: Show History, Compare with Current, Copy Branch Name. The first two route into what already exists — the commit log popover already renders an arbitrary branch's graph, and DiffMode::BranchCompare already backs a three-dot diff — so this is wiring, not a second implementation. Reachable by right-click and, since the picker is keyboard-driven, by the menu key / shift-F10. Compare is disabled on the current branch and navigation steps past it.

Unread panes

The bell indicator only ever came from the shell ringing BEL.

• A manual mark: "Mark as Unread" in the terminal context menu (flipping to "Mark as Read" when a bell is already lit) and a ToggleUnread action on cmd-u / ctrl-shift-u. The render path clears the bell on every frame the pane is focused, which would undo the mark instantly, so a manual_unread flag holds it — set by hand, skipped by the clear, released when focus leaves.
• The mark shows where the pane is hidden: tabs and the minimized/detached chips now report it with the sidebar's glyph and colour, outranking the hook, waiting and active states. That is the case where being told to come back matters most. Toggling also refreshes the windows, since the sidebar sits behind a .cached() wrapper a notify from the pane never reaches.

Wrapped-URL detection

Two follow-ups to the phase-2 guard that extends a URL onto the next row when a TUI wraps the line itself.

• The width guard compared against the wrong thing. "A continuation can never be wider than the row it continues" was given the same +3 slack the surrounding guards had at the time; those lost it later (the layout edge is exact, so slack only lets prose in) and this one was missed. It now compares against the edge itself.
• The width guards cannot decide the narrow case at all: word-wrapping at width W produces rows no wider than W, so a URL row landing exactly on W is the widest row in the block. Read the break lexically instead — an opening paren starts a bracketed token, so it is a parenthetical rather than the URL's tail; and a wholly numeric last segment (/pull/567, /issues/42) can only continue with more digits or with a delimiter opening the next segment. The numeric rule covers PR and issue links without knowing about any particular forge, and still joins a wrap falling inside the number.

Compact header

Revealing the hide/focus buttons on hover widened the cluster to the right of the base-compare chip (branch +N -M), so the chip slid left every time the pointer entered the corner. The buttons move to the git status row, where the flex spacer absorbs their width. The comfortable two-row layout keeps them in the header row — there the chip already sits on a row of its own.

Markdown wrapping

Two layout bugs in the rendered Markdown view, both from min-width: auto on flex children.

• Long frontmatter values ran off the card. The value column of an entry is a flex child, so it was pinned to its unwrapped single-line width and a long description: painted straight past the edge of the metadata card. min-width: 0 lets it shrink and wrap inside it.

• Every inline boundary forced a line break. A paragraph was a wrapping flex row with one item per inline run, so a run long enough to wrap occupied a full-width box and whatever followed it — a code chip, the text after a bold lead-in — started on the next line. This is the limitation Lay out the rendered Markdown as a reading column #185 listed as known and left open.

  Inline elements are now flattened into that row as word tokens: text splits on word boundaries (each token keeping its trailing whitespace), and bold/italic/link hand their emphasis down to the tokens instead of wrapping them in a container. Code stays a single chip, and selection ranges are split per token.

  Measured at 600px against a plain-text baseline of 161px: text interleaved with code chips went from 276px to 161px, a bold lead-in from 184px to 161px. Both shapes are covered by new tests in tests/inline_layout.rs.

Validation

• cargo test on the touched crates — okena-markdown 15, okena-terminal 182, okena-git 175, okena-views-git 35, okena-views-terminal 24, all passing
• cargo clippy --workspace --all-targets and cargo fmt --all --check — clean
• Not done: the Markdown changes were verified by measuring block heights in the test harness, not by looking at a rendered .md in the running app. Somebody should eyeball a document with frontmatter and inline code before merging.

🤖 Generated with Claude Code

https://claude.ai/code/session_01DcL4r46s9HMDmrzQkLzUMo