Give subject the width it needs, and name a commit's branches in the detail panel

[LeadcodeDev]

Three things reported from the same screenshots: branch badges cropped mid-name in the history table, a detail panel that never says which branch a commit belongs to, and a date sharing a row with the author.

Subject takes the leftover width

gpui_component's Column has neither flex nor grow — every width is a number of pixels, resolved once per TableState::refresh. Subject was pinned at 420, so on any window wider than the five columns put together the table stopped short of the right edge, the leftover showed as dead space past Date, and a badge was cropped two columns to its left.

The panel now measures its own bounds through ElementExt::on_prepaint and hands them to the delegate, which subtracts the four fixed columns and the twelve pixels of render_last_empty_col. Subject gets the rest, with a floor of 240 past which the table scrolls horizontally instead.

The delegate answers whether the width actually moved and the panel refreshes only when it did — a refresh notifies, a notify prepaints, and the prepaint is what reports the width, so an unconditional refresh here is a loop, not a measurement. That constraint is now recorded in CLAUDE.md.

A badge renders whole, or is counted

The strip from #49 was a scroller clipped at half the subject cell. It put a branch name cut mid-word in front of the reader and asked them to discover that the fragment could be dragged. Widening Subject made the cut rarer without making it any less of a lie about the name.

Badges that do not fit are dropped whole and replaced by a single +N. Clicking it opens a popover holding the badges it stands for, drawn exactly as the ones on the row and carrying the same context menu — so a local branch pushed out of the strip can still be deleted.

That started as a tooltip listing the names as plain text. A tooltip dismisses the moment the pointer leaves the trigger, so those names could be read and never reached, and a branch reachable only until you move towards it is worse than one plainly absent. badge_cell is shared between the strip and the popover rather than duplicated: the counter changes how many names are on screen, never what can be done with them.

Fitting has to be predicted rather than laid out: gpui hands an element its bounds one phase after its children are built, so the choice of what to build cannot read them. The subject column's width is already known to the delegate, so the strip shapes each name through the window's text system and adds Tag's own xsmall metrics — text_xs, px_1p5, border_1, kept in rems so they follow Window::rem_size rather than a hardcoded pixel count.

fitting_badge_count walks down from "everything fits" rather than up from nothing, because dropping the last badge takes the counter's own reservation away with it: the largest set that fits is not always reachable by adding one badge at a time. A mismeasurement cannot crop anything either way — the budget is half the cell, so a strip that overruns its estimate still lands well inside the cell and costs the subject some room at worst.

One rule then overrides that arithmetic: the first badge always renders, whatever the budget. A row shrunk to a counter alone has lost the one name it had room to say, and a +1 standing for the row's only reference says strictly less than the reference does. This is the sole case where the strip may overrun its budget, and it costs at most one badge.

Suppressing the counter for any single leftover was tried and reverted. Two long names of nearly equal length — feat/x beside origin/feat/x — then overran the whole subject cell rather than merely the budget, and overflow_hidden cut the second one: the crop the counter exists to remove. A +1 beside a name that fits is the lesser loss, and the rule was only ever needed for a row whose only reference does not fit — which the leading-badge rule already covers.

The detail panel names the commit's branches

The General tab now opens with the same reference badges the table draws, above Subject.

For a branch tip that is the same set the table shows. For anything else the answer comes from History::owning_references, which walks down the first-parent chain — from the commit to the children whose first parent it is — until a level carries a reference.

First-parent is what makes the answer read as "the branch this commit was written on". A commit made on a feature branch is the first parent of the next feature commit, and the branch's merge into the trunk takes it as a second parent, so the walk climbs the feature to its tip and stops rather than spilling onto the trunk.

The rejected alternative was every branch that contains the commit, the git branch --contains set. It is the honest answer to a different question and unusable here: an old commit on the trunk is contained by every branch cut since it, which on the repository this was built against is thirty badges above a subject. Stopping at the first level that carries a reference keeps it to the one or two branches that actually claim the commit — a fork point legitimately answers with several.

Walked on demand rather than indexed at load: one pass over commits to build the child map, run when a selection or the history changes, never per frame.

badges moves out of history to the crate root. A branch has to read as one thing across the window, and two copies of the colour mapping would drift apart. The detail panel's strip wraps rather than counting — a header is free to take a second line, where a fixed-height table row is not.

Date on its own row

The author's timestamp shared a row with the name and the address. Widened, that ran past the panel and wrapped, which put the date on a second line with no label beside it. It is a Date row under Author now.

Verification

cargo test --workspace, cargo clippy --workspace --all-targets -- -D warnings and cargo fmt --all --check are green. New tests cover the width arithmetic, the refresh guard, the badge-fitting count (including the case where the counter's own width pushes out the last badge that would otherwise fit), and owning_references over a trunk that merges a feature branch — tip, mid-branch commit, fork point, and a history with no references at all.

The app was launched against a real repository and paints, but the visual check is the reviewer's: this machine has no screen-recording permission, so no before/after capture is attached.

No issue to reference — reported directly in conversation.