chore(UNTRACKED): fix notion bugs

CHANGELOG.md

@@ -1,7 +1,13 @@
-## [1.6.20]
+## [1.6.21]
 
 ### Fixes
 
+- **fix(notion): parse one-way (`single_property`) relation columns instead of aborting the whole database.** `RelationProp.from_dict` only handled two-way (`dual_property`) relations and `raise`d `ValueError("single_property type not recognized")` for anything else. A normal one-way relation column — the Notion default — therefore failed to map, and because the error propagates up to database retrieval it aborted extraction of the *entire* database (every property and row), not just the relation field. `dual_property` is now optional and only parsed when present; one-way relations map with `dual_property=None`, and unmodeled future relation `type`s no longer raise. Adds regression tests for single/dual/unknown relation variants and the relation cell.
+- **harden(notion): drop unknown keys across the page-block extraction path so additive Notion API fields can't break mapping.** Notion ships new, backwards-compatible keys on object payloads (e.g. ordered-list `list_format`/`list_start_index` on `numbered_list_item`), but the connector mapped most types with a bare `cls(**data)` that raises `TypeError` the moment an unmodeled key appears — which had already taken down `numbered_list_item` downloads. A shared `init_from_dict` helper (in the notion `interfaces` module) now instantiates each dataclass from only its declared fields, dropping unknown keys while still surfacing genuinely-missing required fields. It is applied across the block layer (`paragraph`, `heading`, `quote`, `toggle`, `to_do`, `template`, `table_of_contents`, `numbered_list_item`, `child_database`, `child_page`, `embed`, `equation`, `link_preview`, `link_to_page`, `table`, and the `callout` icon sub-types), the rich-text leaves parsed inside nearly every block (`Annotations`, `Text`, `RichText`, the `Mention*` types, `Equation`; `RichText` also tolerates a missing `annotations` key), the file objects backing image/video/file/PDF blocks (`External`, `File` — `expiry_time` is now optional), and the `relation` database property. `link_to_page` variants the dataclass doesn't model (e.g. `comment_id`) now map cleanly and simply render nothing instead of raising. Adds comprehensive unit coverage for every block type plus the helper, rich-text, file, and relation hardening.
+
+## [1.6.20]
+
+### Fixes
 - **test(stager): de-flake the stager bounded-memory tests.** `test_process_whole_peak_memory_is_flat_as_input_grows` and `test_blob_store_stager_peak_memory_is_flat_as_input_grows` asserted a near-flat `tracemalloc` peak ratio between two input sizes. `tracemalloc`'s peak includes not-yet-collected per-element garbage, so the ratio varies with GC timing across environments and the blob-store test failed on CI even though the streamed peak (~3.8 MB) was a small fraction of the ~47 MB input. The tests now assert the meaningful streaming-vs-whole-file bound — peak stays below the input file size (a whole-file load materializes the parsed list at several times the JSON text) — which is robust to GC timing while still catching a regression to whole-file loading.
 
 ## [1.6.19]