[guijacquemet]

Fixes the tables theme of the round-trip findings; their skipped matrix cases and properties now run as normal tests and the findings are removed from tests/roundtrip_known_failures.py. The table output writer is untouched. Code: new _read_table in labconstrictor_tools/convert.py (constants PANDAS_NA_WORDS, NAN_SPELLINGS, EMPTY_CELL).

Before -> after

• F2: floats are read with float_precision="round_trip": 0.30000000000000004 arrives as itself (was 0.3). Cost: about 2x on very large tables (1M rows x 10 float columns: 5.7 s against 2.7 s); documented in PROTOCOL.md.
• F3: duplicate and empty column names arrive as written (a,a -> ["a","a"], ,b -> ["","b"]; before a.1 and Unnamed: 0). The header row is read separately, the body by position and the names put back; a tool addresses such columns by position (documented).
• F4: a missing table file is file_not_found naming the file; an empty or non-UTF-8 file is unreadable_table with an instruction to save a UTF-8 CSV with a header row (traceback logged). Before: raw FileNotFoundError, EmptyDataError, UnicodeDecodeError.
• F7: empty cells are missing; words such as NA, N/A, null, None stay text (a gene called NA, a status None). In a column whose other cells are all numbers they become NaN (a second narrow read). NaN/nan/NAN always mean a missing number: a column of only those spellings (plus empty cells) is float64 NaN, as before. I found and had fixed a regression here before opening this PR: an all-NaN column (what Fiji's ResultsTable, QuPath and numpy/pandas write for an unavailable measurement) had arrived as text. inf/-inf/Infinity columns still come out as floats.

Tests: new tests/test_table_input.py (floats, names, errors, NA words, numeric column with NA, header-only table, all-NaN columns in three spellings, all-NaN plus empty, NaN with numbers, a text column holding nan, an inf column). One existing-test edit: test_roundtrip_known_failures.KnownFailuresDoNotRot.test_the_skip_is_exactly_the_list asserted that an F4 fixture id is a known failure; it now asserts it is not (other fix PRs touch the same line: expect a trivial conflict). Docs: PROTOCOL.md table bullet (states the NaN rule in one sentence), AUTHORING.md, REGRESSION_LEDGER.md (rows moved to Fixed; replace "this PR" with this number).

Host follow-ups: none for reading (hosts only write CSV and send the path); a host writing a table must write UTF-8 with a header row, empty cells for missing, and a text cell such as "NA" is now kept as text; hosts showing unreadable_table or file_not_found should treat them like the image codes.

Verified by the author on Linux, Python 3.11: black, ruff, mypy; the full suite before the last change (only the two environmental sandbox failures; run in groups because a single discover run was killed twice in the sandbox) and, after it, 103 table/properties/known-failure/ledger/unused-code tests and the sabotage test. I re-ran lint and the table tests and checked an all-NaN column (float64), an all-NaN-plus-empty column (float64) and a text column with NA/null (text). Not verified: the full suite and the coverage floor after the last change (CI runs both), Windows, macOS.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Y2SN8SuEakTdiWy3F2WUqN

---

Generated by Claude Code

[guijacquemet]

Closed: included in #47 (merged with a merge commit, so these commits are in main).

---

Generated by Claude Code