[kvlonge]

Why

Cursor-based incremental models had no defined behaviour with no inputs or empty cursor inputs. Empty inputs could fail late after staging work, or leave an existing table empty on a rebuild.

Changes

• Breaking: a cursor-based incremental model that reads no __ref()/__source()/__seed() and lists no cursor_inputs (for example one reading a hard-coded table) is now compile error P011, for every incremental strategy (delete_insert, append, merge). Such models previously re-read everything on every run.
• Empty cursor inputs on an incremental run: the model succeeds without changing its destination, its audits still run, and plan/build show no input rows (<input>.<column>). JSON: empty_cursor_inputs (build), resolution_status: "no_input_rows" and empty_inputs (plan).
• When only some inputs are empty and others have new rows, plan and build warn waiting on empty input <input>.<column>; other inputs have new rows, counted in WARN= and the JSON warnings, with waiting_on_empty_inputs: true. Exit code and status are unchanged.
• A first build over empty inputs creates the destination empty.
• A rebuild that would replace an existing table with nothing (replay_on_change full, aliased models included, or watermark microbatch --full-refresh in any watermark mode) fails with S302 and keeps the rows.
• Models whose inputs are built earlier in the same run follow the same rules. Previously they could advance past an empty input and skip rows it delivered later.
• Inputs are named by project name, including inputs built earlier in the same run.

Verification

• New E2E tests cover empty first builds, emptied incremental runs, the waiting warning (planned and same-run), aliased and microbatch rebuild refusal, project-name labels, P011, and Dagster check results on a no-op run. Text and --json outputs are both checked.
• One independent review plus follow-up; make test, make check-ci.