v1.108.266 - A blank line inside a table cell no longer truncates it

Silent data corruption in the MUNCH decoder. Found in-house, no user report.

The defect

Two functions in encoding/format.py disagreed without saying so. assemble joins payload sections with a blank line and split_sections splits on one, while write_table uses csv.writer, which wraps a cell containing newlines in quotes but keeps the newlines real.

So a cell whose value contains a blank line looks exactly like a section boundary, and the row gets cut in half.

It is worse than truncation, because the row count survives. The orphaned second half becomes its own block, and read_table filters by tag, so the fragment is dropped without comment:

blocks: 2   rows in: 2   rows out: 2
sym_a -> 'def f():'          # was 'def f():\n\n    return 1'
sym_b -> 'def g(): pass'     # intact

Two rows in, two rows out, one cell quietly missing its middle. Nothing raises, nothing warns, and no arity check can catch it.

The trigger is a blank line, not a newline. Multi-line cells already round-trip correctly. What supplies a blank line: a dict or list literal with a blank line between groups, a docstring with a paragraph break, or any span an outliner captures across one. It is latent rather than active only because current encoders mostly capture signatures and short spans. Any future encoder that widens what goes in a cell trips it, and the symptom would be "the outline is subtly wrong sometimes", which is the worst kind to chase.

The fix

split_sections stops treating a blank line as a boundary while the preceding block still has an open RFC 4180 quoted field, and re-joins instead. Detection is a quote-parity count: doubled quotes are the escape form and cancel, so an unterminated field is exactly an odd total.

• No wire-format change. Encoders emit identical bytes; only the reader got stricter about what a boundary is.
• Repairs payloads already written. Anything that hit this decodes correctly now.
• Malformed input degrades instead of vanishing. An unterminated quote through end of payload keeps its text.

Escaping newlines at write time was considered and rejected: unescaping on read would corrupt legitimate code content, since a source cell containing the two literal characters \n would come back as a real newline. That needs a header flag and a version gate. The reader-side repair has none of that exposure.

Tests

tests/encoding/test_format.py goes 68 to 78. Ten new cases: the reported cell, multiple blank lines in one cell, a blank line at the very end of a cell, a cell of only blank lines, embedded quotes alongside a blank line, two cells each carrying one, a scalars block in front, and CRLF. Seven fail before the fix. Three pass on both sides deliberately, as controls: genuine boundaries near quoted values must still split, an unterminated quote must still degrade, and CRLF must keep working.

Full suite 7404 passed, 7 skipped, ruff clean, CI green on Linux and Windows across Python 3.10 to 3.13.