perf(mutation): don't snapshot the whole blob store on every add

[KylinMountain]

Problem

The serial crash-safe add path from #142 lists .openkb/files (the PageIndex blob store) in both the snapshot path set (_snapshot_add_paths) and hardlink_dirs. So every openkb add hardlinks the entire blob store into a rollback backup — one os.link per existing blob, plus the matching unlink on commit — a cost that scales with total KB size, not with the document being added.

Two concrete downsides:

• O(total blobs) per add on the happy path (link + unlink), growing as the KB grows rather than staying O(this doc). Dwarfed by LLM compile on small KBs, but the wrong scaling on large ones.
• Full byte-copy fallback: on a filesystem without hardlink support (cross-device staging, some Windows / OneDrive / Dropbox / network mounts), _hardlink_or_copy falls back to shutil.copy2 — i.e. a full copy of the whole blob store on every add (potentially GBs).
• Cloud import lists .openkb/files too, but never writes a local blob — pure waste.

Fix

The blob store is append-only by {doc_id}: an add only ever creates the new doc's blob, and that name is assigned during indexing — after the snapshot is taken. So instead of snapshotting the whole tree up front:

• Add MutationSnapshot.track_new(paths) — registers paths created after the snapshot with no backup, so both in-process rollback() and crash recovery (recover_pending_journals) remove exactly those paths. It rewrites the active journal, so a crash after the blob lands but before commit still cleans it up.
• add: drop .openkb/files from the eager snapshot + hardlink_dirs; call track_new(files/*/<doc_id>*) right after index_long_document returns a doc_id.
• cloud import: drop .openkb/files from hardlink_dirs (never touched there).

Everything else in the add snapshot is unchanged — hashes.json, pageindex.db*, wiki/*, concepts/entities (still hardlinked, since those are updated in place via atomic temp+replace) all keep their existing backup semantics.

Behavior notes

• Rollback now removes exactly the new doc's <doc_id>{ext} blob and its <doc_id>/images/ subtree.
• If an add creates a brand-new collection dir and then rolls back, the now-empty .openkb/files/<collection>/ dir is left behind. This is harmless — PageIndex add_document does mkdir(exist_ok=True) and create_collection rmtrees before reuse.
• remove doesn't use the mutation snapshot, so it's unaffected.

Tests

Added to tests/test_mutation.py:

• test_track_new_removes_new_blob_on_rollback — rollback removes the new blob + images subtree; a pre-existing blob is untouched (same inode).
• test_track_new_persists_to_journal_for_crash_recovery — track_new persists so recover_pending_journals cleans a crashed add.
• test_snapshot_add_paths_excludes_blob_store — _snapshot_add_paths no longer lists .openkb/files (but still lists hashes.json).

pytest -q: 906 passed (the 4 pre-existing tests/test_url_ingest.py failures are ModuleNotFoundError: trafilatura, unrelated to this change and also failing on main). ruff introduces no new diagnostics in the changed hunks.

https://claude.ai/code/session_018WiFnTo1YW9mtw47Fzir9K

[KylinMountain]

Self-review (max-effort) turned up a regression this PR's first commit introduced — now fixed in 3749648.

Bug: track_new(glob("*/{doc_id}*")) assumed the globbed blob was created by this add. But PageIndex content-dedups (LocalBackend.add_document returns an existing doc_id and writes no new blob when the same content is already indexed). If hashes.json and pageindex.db diverge — e.g. a prior remove whose PageIndex cleanup failed left the row + blob but dropped the hashes.json entry — re-adding that content makes col.add() return the old doc_id, so a subsequent compile failure would roll back and delete another document's blob. The whole-store hardlink snapshot this PR replaced did not have this bug (a dedup-hit blob shares the backup inode → left in place).

Fix: capture the blob set before indexing and register only paths this add actually created (set difference), guarded by if index_result.doc_id. That also neutralizes an unexpected empty doc_id, which would otherwise glob */* and register — then delete — the whole store. Added two long-doc rollback tests (new-blob removed / pre-existing survives / dedup-hit survives); verified the dedup test fails on the pre-fix code.

Known limitation (documented, not fixed): there's a narrow crash window — if the process is hard-killed between the blob landing during indexing and track_new persisting the journal, the blob is orphaned (recovery won't remove it). It's benign (unreferenced; registry is restored on recovery), and a fuller fix would need the doc_id before the snapshot. Calling it out rather than silently accepting it.

Other review findings were minor (layout-coupling glob shares an assumption with remove's delete path; track_new lacks the kb_dir-containment assert snapshot_paths has) — happy to address if you'd like, otherwise leaving as-is.