[asssslay]

csv-generate can drop its configured eof when the final text record crosses a stream read boundary. The same data produces the suffix at a larger read size, so generated files can inconsistently lose their trailing newline or custom terminator. The documented eof option promises trailing characters.

Reproduces on csv-generate 4.6.2 and the parent revision:

import { generate } from "csv-generate/sync";

generate({
  columns: [() => "xx"],
  length: 1,
  highWaterMark: 1,
  eof: "END",
});
// Before: "xx"
// After:  "xxEND"

The overflowing record is emitted immediately. On the next read, finalization sees an empty data buffer and closes without emitting EOF. Treat the pending text-mode EOF as output to flush, even when no record text remains. The existing end-state guard prevents duplicate suffixes; empty output and object-mode records retain their current behavior. After an overflow, the suffix can be emitted as a separate final chunk.

Four new regressions fail before the fix and pass afterward, covering callback, sync, async iteration and Web Stream APIs. The tests also cover custom row delimiters, delayed output, fixed-size boundaries, exactly-once emission, empty output and object mode.

Related to #555, which fixes fixed-size remainder consumption. This also reproduces with the default fixed_size: false after that change. Generated distributions were rebuilt and checked, and are left to the release build process.