[abrichr]

Part of #1671 (the "Matching" items that compare a null as text).

What changed

• rollback.py fresh_presence, restore branch: old and new values compare through a new _same_text. A null equals only a null, so a stored null no longer equals the text "None". A field missing from the row or from the kept old values gives "unknown" instead of matching. A new value counts as null only when it is declared null or the run passed the parameter as None. An unset parameter no longer counts as a null new value.
• qualify/campaign.py RecordReadback.judge: the selector and the field checks use a new _same_field, with the same rule. The field has to be there, and a null equals only a null.
• runtime/effects/onscreen.py _expected_text: when a field_equals declares a value that has no text (a null or unset parameter, or ""), the check is now INDETERMINATE. Before, it fell back to a selector value, which only showed that the record was on screen.
• runtime/effects/_common.py: TYPED_SUBSTRATES now includes db. That is the reward worker's channel for its SQL read (SqlRecordVerifier over SQLite), so a stored number compares by value there, as it already does for sql.

All four changes fail closed or match the existing record_matches rule. None of them lets a write pass that passed before only because of the str() compare.

How it was tested

• New tests/test_null_text_matching.py, which covers both helpers and the db number match. There is also a new test in tests/test_onscreen_verifier.py.
• I ran 8 files locally: the new file, onscreen verifier, rollback, rollback_qualify, both qualification campaign files, business-rule and declarative qualify. Result: 205 passed, 8 errors.
• All 8 errors come from the module fixture in test_qualification_gate_campaign.py. It hit the 600 s pytest-timeout inside cv2.imdecode (vision template match) while the Mac's load average was about 16. I reran one of them alone and it timed out the same way. That fixture runs the benchmark replayer and doesn't touch the changed code, but I couldn't get it to finish locally, so CI has to show it.
• ruff check and ruff format --check are clean on the changed files. mypy is clean on the four changed modules.

Not done here

These remaining #1671 items are left open:

• undo_proposal dropping None selector keys. Fixing it means carrying nulls through _flat_items request matching.
• The ledger items.
• The null-plant in qualification.
• The declared-key check.
• The SQL guard items.

[abrichr]

Independent adversarial review: BLOCK

From: Codex. Reviewed actual head 558d679 in a fresh detached worktree and locked environment.

[P1] A failed mapping is still treated as an explicit null during rollback. In packages/flow/openadapt_flow/rollback.py:205–207, a parameter merely being present in params lets a None resolution enter new_values. A supplied non-null value outside its declared map also resolves to None. For example, value={param: new, map: {known: mapped}} with params={new: unmapped}, kept old note old, and a fresh row with note NULL returns present: the record holds the run's new values. The value did not resolve, so this is unknown, not evidence that the uncertain write landed. A caller can consequently offer compensation on a guessed presence.

An independent probe reproduces this through validated Workflow/UndoStep models, real persisted prior state, and fresh_presence. Require the original supplied parameter to be None before accepting a None resolution as an explicit null; unmapped or otherwise unresolved non-null values must stay unknown. Add the failed-map case alongside explicit-null and missing-parameter controls.

Validation: all eight affected suites pass on Linux, including the campaign fixture that timed out on the Mac: 213 tests passed. Full Flow Ruff/format and mypy pass (591 sources). Seventeen independent behavior probes pass, covering real HTTP qualification reads, null/missing/text selectors and fields, on-screen refusal, and persisted old-null rollback state. The mapped-value case above fails. Synthetic data only, headless execution under the resource gate; no reviewed source or test was changed or pushed.