LGTM, approving. Clean addition that mirrors the existing checkpoint_interval plumbing.

Highlights

• validate_delta_interval faithfully mirrors delta-rs grammar (lowercase interval, closed unit list, non-negative integer) so misconfigured tables fail fast at config-load time instead of silently falling back at log-cleanup time. The rejection test list (INTERVAL, Interval, 30 days, interval -5 days, etc.) is thorough and the rationale is documented in the function doc comment.
• The behavior change has tests on both sides: validate_delta_interval_accepts_* / rejects_malformed lock the grammar, test_log_retention_table_properties proves the properties actually land in the Delta metadata for both unset and set cases, and the Python integration test test_log_retention_properties_land_in_metadata covers the live pipeline path with checkpoint_interval=1 to exercise the cleanup hook surface.
• warn_about_table_property_discrepancies is a nice touch for the existing-table case where delta-rs silently keeps the old config.

Soft non-blocking notes

1. False-positive risk in the discrepancy warning (output.rs, the new check closure): for enable_expired_log_cleanup the user-supplied value is Some(true) (a no-op vs the delta-rs default), but a freshly created table may not write delta.enableExpiredLogCleanup into the metadata configuration at all, so actual.get(key) returns <unset> and the warning fires spuriously. Consider treating <unset> as "matches default" for this specific key, or only warning when the requested value disagrees with the documented Delta Lake default.
2. The warning lives behind mode = append against an existing table — that path is the most user-visible reason these properties get silently ignored. The Rust test test_log_retention_table_properties and the Python test only exercise the "fresh table" happy path; an integration test that pre-creates a table with different delta.logRetentionDuration and then runs the pipeline in mode = append would lock the warning behavior and the silent-ignore contract. Not blocking, but a logical follow-up.
3. Manual Test Plan box says "Tested under various scenarios and modes" — the description boilerplate from #6027 would be more useful if it named the scenarios (fresh truncate, append-existing, append-fresh, invalid-interval). Small ask.
4. Nit: log_retention_duration is a String because delta-rs wants the raw interval string, but it would be more ergonomic for callers to accept a Rust Duration and serialize it to the delta grammar internally — would also remove the need for validate_delta_interval to mirror upstream. Out of scope for this PR; just noting.
5. The new fields are added to the existing Default / make_config test helpers in three benches plus the parallel-test fixtures with : None, — good, that catches any future struct literal that forgets them.

PR description still has the Changelog updated checkbox unchecked; please tick it (or add a changelog line) before merge if you have a changelog rotation. Nothing else.