Keep filtered lists fresh; derive observed status per sky object

[brickbots]

Summary

Follow-up to #526: the filter cache keys all reuse on dirty_time, which only advances when a filter parameter changes — but object state changes under an unchanged filter too. This PR keeps filtered lists correct in those cases, and makes observed status behave as a property of the sky object rather than of the catalog listing.

Four commits, one concern each:

1. fix(catalogs): keep filtered lists fresh as objects are logged and the sky rotates

   • Logging an object invalidates the cache when an observed criterion is active (Catalogs.mark_logged), so an "Observed: No" list drops it on its next refresh.
   • With an altitude criterion active, verdicts age out (CatalogFilter.is_stale(), TTL 600 s ≈ 2.5° of sky drift, inside the 10°-step filter granularity), and the arrival of an alt/az fix triggers one refilter — closing the boot-time gap where pre-lock verdicts skipped the altitude test and everything "passed" forever. Staleness is promoted to a dirty bump in filter_catalogs() so both cache layers re-evaluate, and UIObjectList.update() polls it so a list screen left open refreshes in place.
   • Without an altitude criterion is_stale() short-circuits — perf(catalogs): skip re-filtering unchanged catalogs on list open #526's O(catalogs) list-open fast path is preserved.

2. feat(observations): derive observed status per sky object, not per listing
   Logging M 31 marks NGC 224 observed — in-session, across restarts, and retroactively for historical logs. Log entries stay recorded per (catalog, sequence) (no schema change); the observed cache additionally maps each logged listing to its object_id at load, check_logged tests id membership for DB-backed objects, and mark_logged propagates to sibling composites. Virtual objects (planets/comets/coordinate entries with session-minted negative ids) keep the per-listing test — logging Mars must not mark Jupiter.

3. feat(observations): show log entries from all sibling listings on details
   get_logs_for_object resolves a DB-backed object's sibling listings, so NGC 224's details show M 31's logs ("1 Logs") instead of contradicting the row's checkmark.

4. fix(ui): keep the cursor on target across object-list refreshes
   Refreshes re-locate the previously selected object; if it was filtered out (just logged, or dropped below the altitude criterion), the cursor falls to the first surviving old successor — the natural next target — instead of resetting to the top. Matching is by (catalog_code, sequence), since CompositeObject.__eq__ compares object_id alone and would land on a sibling listing.

User-visible after upgrade (intended): historical observations retroactively mark sibling listings — an "Observed: No" NGC list may visibly shrink, sibling rows gain the observed checkmark, and a sibling's details show the combined log count.

Docs

• ADR 0020 (docs/adr/0020-filter-freshness-staleness-promotion.md) records the freshness policy and the observed-identity decision, including rejected alternatives (timer thread, per-object invalidation, in-memory-only propagation, schema migration).
• docs/ax/catalog/CONTEXT.md: Logged redefined per sky object, Log entry and Stale terms added; docs/ax/catalog.md §4 updated to match.

Test plan

• 14 new unit tests across test_catalog_filter_cache.py, test_observed_identity.py, test_object_list_cursor.py (freshness triggers, sibling propagation, restart derivation, virtual-id fallback, sibling details logs, cursor next-target rule).
• Full suite: 803 smoke+unit pass on top of current main; ruff + mypy clean.
• Verified live (headless app) against a real observations DB: sibling checkmarks derived purely from historical logs (NGC 224/221/225 via M 31/M 32/Col 7 entries, unlogged neighbors unaffected), "1 Logs" on sibling details, a just-logged object dropping from an "Observed: No" list with the cursor landing on its successor, and the altitude filter applying live.

🤖 Generated with Claude Code