CLN: correct issue refs and comments on the read_csv token-length helper

Comment-only follow-ups from reviewing the branch:

- The embedded-NUL comment and test cited GH#66277, the perf issue, and
  briefly GH#19886, which this branch does not fix (three of that issue's
  four reported cases still fail). Both now cite GH#66415, matching the
  whatsnew entry for the fix.
- The helper's precondition named coliter_next; its only caller uses
  coliter_next_with_idx, so the precondition is now checkable.
- The test comment explained the wrong thing about the two columns: "a"
  covers the words[token_idx + 1] branch and "b" the stream-end branch.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Author: jbrockmendel

pandas/_libs/parsers.pyx

@@ -2571,7 +2571,7 @@ cdef _string_pyarrow_utf8(parser_t *parser, int64_t col,
                 continue
 
             # Not strlen: an embedded NUL is a data byte here, so strlen would
-            # truncate the field at it (GH#66277).
+            # truncate the field at it (GH#66415).
             wlen = _token_len_words(parser, token_idx, word)
 
             if not large and total_bytes + wlen > <Py_ssize_t>INT32_MAX:
@@ -3062,11 +3062,11 @@ cdef inline int64_t _token_len_words(parser_t *parser, int64_t token_idx,
     # rather than `word_starts`.  The tokenizer keeps the two in lockstep
     # (words[i] == stream + word_starts[i], rebased whenever the stream
     # reallocs), so the result is identical; what differs is which array the
-    # loop touches.  coliter_next already loaded words[token_idx] to produce
-    # `word`, so the boundary comes off a cache line the loop has in hand
-    # instead of streaming a second metadata array alongside the first.
-    # `word` must be the unmodified coliter_next result for `token_idx`;
-    # an adjusted pointer would silently yield a wrong length.
+    # loop touches.  coliter_next_with_idx already loaded words[token_idx] to
+    # produce `word`, so the boundary comes off a cache line the loop has in
+    # hand instead of streaming a second metadata array alongside the first.
+    # `word` must be the unmodified coliter_next_with_idx result for
+    # `token_idx`; an adjusted pointer would silently yield a wrong length.
     # Only the pyarrow string path uses this: rewriting _token_len itself to
     # this form regressed long-token ints ~6% (GH#66277), so the numeric
     # callers keep the word_starts version.

pandas/tests/io/parser/test_c_parser_only.py

@@ -858,10 +858,12 @@ def test_block_lane_nrows_short_row_near_stream_capacity(c_parser_only):
 @pytest.mark.parametrize("kwargs", [{}, {"dtype_backend": "pyarrow"}])
 @pytest.mark.parametrize("prefix_len", [1, 200])
 def test_embedded_nul_byte_roundtrip(c_parser_only, kwargs, prefix_len):
-    # GH#19886: the pyarrow string fast path computed token lengths with
+    # GH#66415: the pyarrow string fast path computed token lengths with
     # strlen, so a quoted field with an embedded NUL byte was truncated at the
-    # NUL.  Column "b" is the last token in the stream, so its length comes
-    # from the stream end rather than from the next token's start.
+    # NUL.  Column "a" takes its length from the next token's start and column
+    # "b", last in the stream, from the stream end, covering both branches of
+    # the length helper; the two prefix_len values keep a length-capped scan
+    # from passing by accident.
     pytest.importorskip("pyarrow")
     parser = c_parser_only
     value = b"x" * prefix_len + b"\x00y"