polish(errors): actionable delete-manifest rejection message

[gordonmurray]

The previous 400 response told users what went wrong but not how to fix it:

Table contains row-level deletes which are not supported. This application only supports append-only Iceberg v1/v2 tables. Reading this table may return incorrect data.

The fix is always the same — compact the table. Spelling it out removes a round-trip to Google / the Iceberg docs:

Table has row-level deletes (position or equality). Cloudfloe reads would return rows the Iceberg metadata has marked removed, so the query is blocked rather than silently wrong. Compact the table first — Spark: `CALL system.rewrite_data_files('..

')`; Trino: `ALTER TABLE EXECUTE optimize`; Iceberg CLI: `iceberg rewrite_data_files`. Then retry.

Why keep the hard block

DX.md suggested letting users proceed with a "stale data warning." Rejected — Cloudfloe's positioning (README calls out "direct Parquet reads can return deleted rows" as a danger) depends on not doing that. A toggle that silently re-enables wrong-row behaviour undermines the "reads Iceberg correctly" promise. Better error + hard block beats looser block + silent wrongness.

Verified

• `pytest` 40 passing (ignoring `test_metrics.py` driving feat(metrics): replace placeholder bytesScanned with real DuckDB profile #25, unrelated).
• `ruff check` clean.
• One-line string change; no behaviour change beyond the message.