test(engine): cover camelCase @unique on append and merge loads

[pronskiy]

What & why

#765 reported that writes into a type with a camelCase @unique property failed with No field named versiontag, for load --mode append, load --mode merge and mutate … insert. #768 fixed the committed-uniqueness probe (col → ident in CommittedState::unique_holders), but its regression case, issue_765_camelcase_unique_property_insert.gqt, only covers insert. The two load modes reach the same probe through CommittedState::load, and no test covers them. A GQT case can't: GQT seeds only with Overwrite, which skips the committed probe.

This adds load_issue_765_camelcase_unique_column_loads_and_rejects_duplicates to crates/omnigraph/tests/validators.rs. For both Append and Merge, it loads Message { slug @key, messageId @unique } into an empty graph. It then checks that a second load with the same messageId is refused with @unique violation on Message.messageId and leaves one row.

Backing issue / RFC

• Fixes an accepted issue: Refs bug: any insert into a type with a camelCase @unique property fails — uniqueness lookup sends the column name unquoted (No field named versiontag) #765 (accepted, closed by fix(engine): use literal identifiers for property names #768; this adds the missing load-path regression test)
• Is an RFC PR, or implements an accepted RFC
• Trivial fast-lane

Checklist

• Change is focused (one logical change)
• Tests added/updated for behavior changes
• Public docs updated if user-facing surface changed: N/A, test-only; no changelog note
• Reviewed against docs/dev/invariants.md: N/A, no production code changes

Local verification

• cargo test --locked -p omnigraph-engine --test validators issue_765: 1 passed
• cargo test --locked -p omnigraph-engine --test validators: 35 passed
• cargo fmt --all --check: clean
• cargo clippy -p omnigraph-engine --all-targets --locked -- -D warnings -W clippy::dbg_macro: clean
• Reverting fix(engine): use literal identifiers for property names #768's ident back to col in CommittedState::unique_holders makes the test fail with the original No field named messageid … lance-datafusion-11.0.0/src/planner.rs:1025:20. With the mode order swapped, the Merge load fails the same way, so each mode catches the bug on its own.

Notes for reviewers

• The test loops over the two modes instead of adding one test per mode. Both modes build the same CommittedState::load view, and the loop keeps a single schema and body.
• A camelCase @key needs no test here: keyed loads upsert on the __id system column (MergeInsertBuilder::try_new(ds, vec![system_columns.id])), so the key column's name never reaches a filter.

[azimafroozeh]

Thanks!