[samuelduchesne]

Closes #183

Problem

For an object type with no extensible group, the IDF parser stored every blank field as "" (numeric fields too). write_idf and write_epjson both drop those blanks, so re-parsing idfkit's own IDF output did not equal the original. For example, Zone, Z1, 0, 0, 0, 0, 1, 1, , , , , ; had 12 keys in to_dict() before the write and 7 after.

Change

IDFParser._build_data_dict_cached now skips blank values instead of storing "", for the fixed fields of every object type. This matches what the extensible-group path already did after #182, and what idfkit-js does: packages/core/src/parse/idf.ts skips blank fields and drops trailing blank padding. The loop now zips the named fields, so it no longer walks the extensible tail.

IDFObject._set_field now treats None and "" as clearing a field, as assign_slot already does for an extensible slot. Writing a blank over an unset field does nothing. Without this, obj.f = obj.f on a parsed blank field stored None, bumped the mutation version and cleared the preserved source text. That made write_idf(preserve_formatting=True) reformat an object nobody had changed. On main, the stored "" matched and the assignment did nothing.

Effects:

• Field access: a blank field now reads as None through attribute, item (obj["..."], obj[i]) and get() access. That is what a field cut off by an earlier semicolon has always returned. Under strict=True, an unset schema field still reads as None and never raises.
• eppy compat: fieldvalues is built from field_order with data.get, so a blank slot now gives None, the same as a field never given. It used to give "".
• Writers: no change. write_idf already writes a missing field as an empty slot, so an interior blank keeps its position (Zone,Gap Zone,,0,0,0,,2;). When an edited object is written with preserve_formatting=True, its blank commas are still kept, because of the existing keep_at_least rule.
• An explicit 0 (0, 0.0, -0.0, 0.0E+0) is still a value.

The change is user-visible, since "" became None. So it comes with a CHANGELOG ### Fixed entry and a new "Blank fields are unset" section in the parsing-idf-epjson agent reference, with the bundle re-baked.

Not changed here: doc.add(..., field="") still stores "", because that path is in document.py. A strict parse still refuses a non-extensible object with extra trailing blank fields, as before.

Levels

[tool.idfkit.conformance] level moves from conformance-2026.19 (set by #236) to conformance-2026.20. That level differs from .19 only in removing the two Python round-trip entries this PR closes (idfkit/idfkit-conformance#47). idfkit.CONFORMANCE_LEVEL and the EPW sentinel table are regenerated from the tag; only their recorded level changes.

Tests

tests/test_idf_roundtrip.py::TestBlankFieldsAreUnset covers:

• the truncated and blank-comma spellings giving equal models
• interior blanks being absent, and zeros being kept
• access returning None, in strict mode and through eppy fieldvalues
• write_idf → re-parse giving an equal model, with the same key order
• interior blanks being written in position
• write_epjson equalling obj.data
• the preserving writer keeping the blank commas of an edited object
• re-assigning a blank field not counting as an edit
• assigning None or "" clearing a field

make check passes. make test has one failure: test_weather_offline.py::test_a_read_only_cache_stops_the_nudge_without_the_env_var. It fails the same way on unmodified main in this environment, because the container runs as root and a chmod 0o500 cache directory is still writable for root. It is unrelated to this change.

Conformance

python3 runners/run.py --library ../idfkit at the conformance-2026.20 tag: PASS, exit 0. 74 cases, 226 assertions: 218 passed, 5 failed, all 5 of them outstanding exceptions (#186, #187). No stale entries and no unexpected failures.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LEegXpWqwRPaofngRx6L3h

[samuelduchesne]

corpus (python) is red because the fix works. At the pinned level (conformance-2026.16), the runner reports the two #183 allowlist entries as stale and exits 1:

stale exceptions: these cases now pass, so remove their entries (exit 1)
  numeric-zero-against-blank: python: round-trip now passes
  positional-trailing-unset-no-extensible: python: round-trip now passes

Nothing else changed: the 5 outstanding exceptions (#186, #187) are the same ones main reports.

This can't go green from this repo alone. As with #226 → conformance-2026.16, the unblock is:

1. Remove the two python / round-trip entries for The document model keeps blank trailing fields that both writers strip, so the round trip is lossy #183 from known-divergence.toml in idfkit-conformance, and cut a new conformance-2026.N tag.
2. Bump [tool.idfkit.conformance] level in pyproject.toml here to that tag. I'll push that as soon as the tag exists.

Moving to an existing later tag doesn't help: conformance-2026.19 still carries both entries.

---

Generated by Claude Code