Clean scope: one-commit fork bump + a real regression test. The fork's v0.32.3-feldera-fix branch is rust-v0.32.3 + the two existing patches + the one new patch 297f2dd1 (comparison is 3-ahead / 0-behind of upstream, 1-ahead / 0-behind of the previously-pinned 5b35c098), so nothing is silently rebased or dropped. The Cargo.toml comment tracks all three patches, points at the new compare URL, and the lockfile update matches.

The regression test earns its keep: after is a six-field nested struct whose children are themselves column-mapped to diverging col-<id> names, so it exercises field-id resolution inside the struct rather than just at the top level — that is the "no field name overlap" struct-cast failure mode from #6557, not only the col-102 is missing case. op VARCHAR NOT NULL also covers the hard-error branch, and the fixture is built by pyarrow (via ensure_delta_spark_fixture), which is the only way to get logical-name Parquet on disk. FIXTURE_VERSION is present for cache invalidation. Good hygiene.

Nits, non-blocking:

• Consider marking the test with the same @pytest.mark. used by other delta-input suites so it participates in the same selectors — I only see the bare def test_.... If the rest of test_delta_input_*.py similarly has no marker, ignore.
• Comment nit: # (logical child name, field id) for the six-field after struct. — with AFTER_CHILDREN[0][1] == 104, the top-level ids are id=100/after=101/op=102 and children start at 104 (skipping 103). Not wrong, just worth a one-line note in the fixture header so a future reader doesn't assume 103 is missing by accident.
• Changelog checkbox is unchecked in the PR body; a one-line "delta connector: read UC-Uniform-over-Iceberg column-mapped tables by Parquet field id (#6557)" under Unreleased would be nice for release notes but isn't strictly needed for a bug fix.

Approve. Solid, minimal, tests the actual failure surface, and the fork diff is auditable.

Re-review at tip ee4f35aa.

Additive delta since my last APPROVE (23acfe97):

1. Fork rev bumped 5b35c098 -> 297f2dd1 on branch v0.32.3-feldera-fix, adding a third patch 9a58fbef "delta-rs: resolve column-mapped Uniform/Iceberg columns by Parquet field-id". Fixes col-<id> is missing from the physical schema when reading UC-Uniform-over-Iceberg tables where columnMapping.mode=id. Cargo.toml header updated to list all three patches with the correct compare URL; matches Gerd's convention for the fork-pinning block.
2. New pyarrow-driven fixture python/tests/platform/fixtures/uniform_iceberg.py (217 lines) that hand-writes the Parquet + _delta_log in the exact shape a native Iceberg writer + Unity Catalog Uniform produces: Parquet columns named by their logical names carrying PARQUET:field_id, Delta log using columnMapping.mode='id' with divergent col-<N> physical names — at both top level AND inside the after struct's six children. Delta Spark can't produce this shape, so the pyarrow route is correct. store_schema=False is right to keep the file Iceberg-shaped.
3. New python/tests/platform/test_delta_input_uniform_iceberg.py (139 lines) exercising the customer's cdc_raw failure on two fronts:
   • Non-nullable top-level op (physical col-102) — pre-fix errors with "Non-nullable column 'col-102' is missing from the physical schema".
   • Nested after struct whose six children are each mapped to col-<id> — pre-fix fails the struct cast ("Cannot cast struct with 6 fields to 6 fields because there is no field name overlap"). The test projects a nested child of after, so field-id resolution is exercised at every level, not just the top.
     FIXTURE_VERSION = "v1" bump comment covers the caching gotcha.

QA per swanandx's comment above. Re-approving.

One small ask (non-blocking, per mihaibudiu's L115 comment on my last pass): please do submit the three patches upstream when you get cycles — the fork's diverging longer than we'd like.