[abrichr]

What

On pixel surfaces (RDP, Citrix, canvas), the right-record check abstains when a record ID has a look-alike glyph (1/l, 0/O). Before this PR the run then stopped. Now it takes extra looks first, and stops only if they don't settle it:

• the same OCR read on 2 fresh screenshots (each must read the same look-alike identifier);
• the existing 2x re-read (must not name another record);
• the known-ID list, when configured (already shipped);
• where the surface allows it, a select-and-copy of the identifier's own text box, read from the clipboard (runtime.identity_copy_look in deployment.yaml, off by default).

Each look is recorded on the step's identity check (IdentityCheck.looks).

Fail-closed rules

• OCR looks never accept on their own: a look-alike twin can read the same on every frame and at every size.
• Only the known-ID list, or a copied ID that equals a run parameter in the identity band, can settle the record.
• A recorded band is OCR text and can be a misread (A1001 recorded as Alo01), so in recorded-record mode the copy can stop a run but never accept one.
• Any look that names another record is final (mismatch).
• The copy drags across only the one text box holding the identifier. If the screen changes anywhere else during the look, the run halts (look_halt).
• The copy look is input, so only the step-entry check takes it. The check just before the click takes no input; it reuses the copy only when the identifier's pixels are unchanged.
• Governed and qualification runs never take the copy look.
• A run-parameter mismatch that is only a look-alike misread (C0X1234 read as COX1234) gets the same looks; the mismatch stands unless the list or an exact copy of the run value settles it. It is never softened into an abstain.
• A split ID (103 045 against 103045) is matched as one string, against the read and the list.

Tests

• tests/test_identity_looks.py: decision rule, copy look, crop matching (pure).
• tests/test_identity_looks_replayer.py: real render and real OCR on the replayer path. Correct row verifies with copy + run parameter; look-alike twin AC5OO61 stops; recorded mode cannot accept; a screen change outside the cell halts; governed runs take no copy; the pre-click check reuses the copy only on matching pixels.
• tests/test_known_ids_replayer.py updated: the list now needs 3 agreeing frames instead of 2.
• Local run of 43 identity/deployment/replayer test files: 1278 passed. The 2 failures (test_pixel_identity_probe) are a local scikit-image/numpy ABI mismatch and also fail on main.

Measurement

python -m openadapt_flow.validation.identity_looks_dense drives the real replayer against a live browser render of the dense record list, with real mouse drags. On 288 correct-row clicks, false stops went from 232 to 24 with the copy look and the ID as a run parameter, and wrong-record accepts stayed 0 of 288 in every configuration. Numbers and docs: #1340.

No new commands.

🤖 Generated with Claude Code