Add a Rebuild Index button to the Album Manager

[lstein]

A red Rebuild Index under Update Index, confirmed before it does anything, shown only for albums that already have an index and hidden while one is being built.

It exists because Update Index cannot do this job. An update is a set difference on paths — it adds files that appeared and drops files that vanished — and never re-reads a file already in the index, whatever its mtime. So anything derived at index time is fixed at that moment, and until now the only way to revisit it was to delete the .npz by hand. (That is what made #401's new sidecar metadata invisible on an existing album.)

The flow reuses the path that already recovers from a corrupted index: remove the index, then index the album. Its flag is renamed isCorrupted → removeExistingIndex, since a deliberate rebuild is now the second caller, and the messages along it no longer say "corrupted".

Two bugs found by driving a real browser, not jsdom

The confirmation dialog rendered behind the Album Manager. #confirmModal was z-index: 9999 against the overlay's 99999, so the prompt painted but elementFromPoint over its Cancel button returned the album card underneath — unclickable, flow dead. It is now above the overlay and the file tree, still below the spinner and toasts. A test pins the ordering by parsing the CSS, because jsdom does no layout and could never have caught it. #deleteConfirmModal is the same shape but is only ever opened from the main UI, so it is left alone and noted in place.

rebuildIndex captured the card element before awaiting the confirmation — an await of unbounded length — while loadAlbums() rebuilds the card list wholesale. Progress could be painted onto a detached node with the visible card frozen. It now re-resolves through _liveCardFor, the helper that exists for exactly this.

Neither is visible in a screenshot; the first only surfaced because Playwright could not click Cancel, which was easy to dismiss as a test artefact.

Backend

remove_index now clears the npz lru_cache. Every other clear sits in a write path, which was sufficient while the endpoint only ran as a prelude to re-indexing — the rebuild's own write cleared it moments later. A button makes that rebuild skippable, cancellable and failable, so the app could go on serving an index that is no longer on disk. A test fails without the fix.

Only embeddings.npz is removed. The semantic map, cluster labels and thumbnails are derived and self-invalidate by mtime — umap.npz against the index, the labels against umap.npz — so they regenerate, and deleting them would only throw away reusable work. Pinned by a test.

Layout

The four buttons did not form a grid: .index-controls centred its row against a status message that wraps to three lines, so Update Index floated down while Edit sat at the top of the next cell, and Rebuild landed 9px off Delete. The near-alignment of the first pair was coincidence. The index buttons are now a column mirroring .action-buttons.vertical. Measured across every card afterwards: 0px offset for both pairs, and the two index buttons share a width the way Edit and Delete do.

Testing

1097 backend + 906 frontend tests pass; ruff, eslint and prettier clean. Verified against a real 13-album install: placement, colour, the dialog's wording and buttons, and hide/show across the indexing lifecycle. The confirmation was exercised with Cancel only — no index was destroyed in the making of this PR.

Docs: a new "Rebuilding an index from scratch" section in docs/user-guide/albums.md, which explains why an update cannot substitute. #401 still tells people to delete the index file by hand; whichever of the two lands second should point at this button instead.

🤖 Generated with Claude Code