fix(query): built-in index dispatch fallback — no more idx->> seq-scans (#154)

[jensens]

Summary

Closes #154. On aaf-6 prod a folder-listing query produced 4-9 second seq-scans because `_process_index` fell through to `_handle_field` when the `IndexRegistry` didn't carry the queried index name. For Plone-built-in names (`path`, `allowedRolesAndUsers`, `effectiveRange`, `SearchableText`, …) the generic `idx->>'name'` SQL bypasses the dedicated typed columns — `path` column, `allowed_roles` GIN array, `object_provides` GIN array, `searchable_text` tsvector, and the `effective`/`expires` composite clause — and their indexes.

Observed SQL (slow, 9 seconds, seq-scan):
```sql
WHERE idx IS NOT NULL
AND idx->>'portal_type' = ANY(...)
AND idx->>'path' = ANY(...) -- should use path column
AND idx->>'allowedRolesAndUsers' = ANY(...) -- should use allowed_roles && ...
AND idx->>'effectiveRange' = -- nonsense scalar equality
```

Fix

`_process_index` now calls `_builtin_index_type(name)` on registry miss:

• Hardcoded Plone-native specials: `path` → PATH, `effectiveRange` → DATE_RANGE, `SearchableText` → TEXT. Their SQL lives inside the handlers (composite clauses, tsquery matching) so they can't be expressed via `ExtraIdxColumn`.
• Derived from `ExtraIdxColumn`: every `TEXT[]`-typed column entry gets auto-registered as a KEYWORD keyword-like index. Currently `allowedRolesAndUsers` → `allowed_roles`, `object_provides` → `object_provides`. Addons that call `register_extra_idx_column(col)` with `column_type="TEXT[]"` get the fallback treatment for free.

Explicit registry entries still win so addons can override handler behavior.

Test plan

• 7 new tests in `TestBuiltinIndexFallbackDispatch`:
  • `path` unregistered → uses `path` column, not `idx->>'path'`
  • `allowedRolesAndUsers` unregistered → uses `allowed_roles && %s::text[]`
  • `effectiveRange` unregistered → composite `effective/expires` clause
  • `SearchableText` unregistered → `searchable_text @@ tsquery`
  • `object_provides` unregistered → TEXT[] `&&` (proves the ExtraIdxColumn derivation works)
  • Explicit registry entry still wins (regression guard)
  • Truly custom/unknown index still falls through to `_handle_field` (addon-defined)
• Full suite: 1450 passed / 23 skipped / 0 failed.
• After deploy on aaf-prod: EXPLAIN on a folder-view query uses `Index Scan` on `idx_os_cat_path` and `idx_os_allowed_roles`, not `Seq Scan`. Magazin-view latency drops from 4-9 s to sub-100 ms.

🤖 Generated with Claude Code