librarian: mount + seed runtime JSON state under /lib_temp_files

(Re-applied cleanly on top of main: PR #18 was rebase-merged, so the branch
that became #19 conflicted on the already-landed commits. This carries ONLY
the librarian state fix - the sole content difference between that branch
and main - so nothing else is touched or lost.)

The worker threads open cr_results/rr_results/not_in_db/s_results.json in
place ('r+'), crashing with FileNotFoundError in a fresh container. New
conjurer_librarian/lib_paths.py resolves all four under a persistent dir
(CONJURER_LIBRARIAN_STATE_DIR, default /lib_temp_files) and seeds missing
ones with '{}' on import; shared by conjurer_librarian.py and scrape_bot.py
(no circular import). ndb_database/database initialised to {} before load so
a corrupt persisted file degrades to empty instead of NameError. search_bot
/search_bot2 open the DOI DB 'r' not 'r+' so /doi can stay read-only. Docker:
STATE_DIR env + VOLUME, compose mounts /srv/librarian/state:/lib_temp_files.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

docker/compose.librarian.yaml

@@ -14,7 +14,11 @@ services:
     ports:
       - "5001:5001"
     volumes:
-      # Local DOI chunk database (0_chunk.txt ... N_chunk.txt).
+      # Local DOI chunk database (0_chunk.txt ... N_chunk.txt). Read-only:
+      # the search workers only read it (open mode "r").
       - /srv/librarian/doi:/doi:ro
+      # Runtime JSON state (cr_results/rr_results/not_in_db/s_results) -
+      # seeded on first run, persisted here across restarts.
+      - /srv/librarian/state:/lib_temp_files
       # Optional: netrc holding Crossref credentials (or use CONJURER_CROSSREF_MAILTO).
       - /srv/librarian/secrets/.netrc:/secrets/.netrc:ro