fix(ui): keep what is loaded on screen while it reloads

[LeadcodeDev]

Deleting a local branch made the history blank out twice, about a third of a second apart.

Why it blanks

reload_head, reload_references and reload_history each assigned LoadState::Loading before spawning their read, discarding the value the views were drawing. HistoryTableDelegate::recompute_visible returns early on anything but Ready, so rows_count drops to zero and the table renders nothing until the walk lands. The sidebar loses its branch list the same way.

So any reload flickers — a commit made in a terminal, a fetch, a scope change. Deleting a branch is just where it shows up twice.

Why twice

Nothing deduplicates the two reloads a deletion causes:

1. delete_local_branch reloads the moment git branch -D returns (workspace.rs:619).
2. 200–350 ms later FSEvents reports refs/heads/<branch> gone, classify maps it to Aspect::References, and the watch loop reloads again.

The gap is FSEvents latency + the 200 ms debounce window + up to one 100 ms mailbox poll — long enough for the two blanks to read as separate flashes.

I confirmed the echo rather than assuming it: watching a repository through FsRepositoryWatcher across a git branch -D reports Aspect::References every time.

The change

LoadState::start_reload moves to Loading only when there is nothing worth keeping. A first load and a retry after a failure still show the loading state; a re-read leaves the previous value on screen until apply_* replaces it.

One change, three call sites, and it fixes the flicker for every reload path rather than just deletion.

It also fixes a smaller artifact in the same family: apply_history reads self.references.ready(), so when references was still Loading it built an empty refs_by_commit and every badge vanished for a frame.

What I deliberately did not do

The redundant second reload stays. It is invisible now, and suppressing gitr's own echo would need the watcher to tell gitr's writes from a terminal's — domain::watch leaves that to the caller by design, and nothing in the tree does it. Worth its own change if the duplicate walk ever costs measurably on a large history.

Testing

Four unit tests on the state transition, written before the code and seen to fail. cargo test --workspace, cargo clippy --workspace --all-targets -- -D warnings and cargo fmt --all --check all pass.

Not yet confirmed in the running app. gitr is single-instance and the installed binary held the socket, so I never got the fixed build in front of a human eye. The reasoning above is from reading the code plus the watcher probe — the visual result is unverified.