fix: rename extensions, watcher path prefixes and scribe queue ordering

[ajslater]

Quick fixes from an audit of the rename/move pipeline (the work behind #822 / #823 / 7c15ff221). Each is independent of the others; none of them change the rename architecture, which is a later change.

The audit itself found the recent rename PRs sound — every codex-initiated rename event traces end-to-end to a preserved row and bookmarks. These are adjacent defects it turned up along the way.

Fixes

Rename gave every archive a .cbz name. ext is a comicbox metadata field rather than the file's suffix, and codex's read config deletes it, so comicfn2dict fell back to its "cbz" default: every PDF and unconverted CBR/CBT/CB7 was renamed to a name claiming to be a zip, and the admin preview showed the same wrong name. Renaming now uses a config that keeps ext and states the archive's real suffix as metadata. Both halves are required — un-deleting the key leaves it unset, and delete_keys is applied after the merge, so it would strip a stated value under the read config. comicbox still performs the rename; it just gets correct inputs. Stating the extension also outranks one a third-party tagger embedded in the archive.

Deleting a watched folder deleted sibling folders' comics. expand_dir_deleted matched a bare path prefix, so deleting /c/Batman collected everything under /c/Batman Beyond too. Those comics were deleted with no paired add to rescue them, cascading their bookmarks away while the files were still on disk. Prefixes now terminate at a separator.

Sibling libraries stole each other's events. _find_library had the same bare-prefix bug, so with roots /comics and /comics-kids (a pair the admin serializer permits — it only rejects nesting) the second library's events were filed under the first, importing comics at paths outside their own library.

Two TypeErrors in the scribe priority queue. SHUTDOWN_MSG was a bare int where real items are (priority, timestamp) tuples, so stopping the thread with any task queued raised comparing int to tuple and aborted the daemon's shutdown loop before the remaining threads were told to stop. Separately, equal priorities fell through to comparing ScribeTask dataclasses, which define no ordering — a lost task in the routing thread. Timestamps tie more readily than they look (truncated, and clocks can step backwards), so a monotonic counter now closes the tuple.

Online-tag merge never deduped. _merge_task tested candidate paths against path_to_pk's values, which are pks — a Path never equals an int, so an overlapping second scan re-queued every shared comic: inflated totals and duplicate lookups against rate-limited sources.

The settle-timeout warning was misleading. It told the admin to poll again once copying finished, implying the task was abandoned, while the import ran on regardless — and skipped starting its statuses on the way out.

Notes for review

• The settle timeout keeps importing rather than stopping to match its old message. Stopping would drop those events entirely on a watched library that isn't also polled; the current behavior is self-healing, since files still mid-copy land in FailedImport and are retried.
• COMICBOX_RENAME_CONFIG is separate from COMICBOX_CONFIG so read/import parsing is untouched — this only widens what the rename pass parses.
• New tests use a real archive. Whether the rendered extension is right now depends on what codex hands comicbox, which the existing test double can't exercise, so TagWriterRenameExtensionTests repacks the example CBZ as a CBT and runs real comicbox — including one test asserting the admin preview equals what the rename produces. Every new test was verified to fail against the old code.
• make lint, make ty, full pytest and vitest all pass.