[abrichr]

Closes #1722

What changed

• column_roles in generate/demo_analysis.py used to label a column "holds the row's own key" before it checked for a copy. Now that label applies only after _copied finds no copy. A copied price of 15 and 25 on rows 1 and 2 reads as {key}5, but it is now copied, and the contract checks it. A column that really holds its own key (AUD-{id}) is still listed with that reason in the question.
• _key_template became _holds_own_key, which returns a bool. Nothing read detail["template"].
• In docs/GENERATE_CONTRACTS.md, the own-key paragraph now describes what the classification does: it is a reason shown in the question, the column is left unchecked only if the answer names it, and a copy is checked as a copy even when its values read as the own key.
• tests/test_generate_coincidences_r1.py:
  • The fixture no longer names sales.price under unchecked.
  • test_a_value_that_reads_as_the_own_key_is_asked_never_templated is replaced by test_a_copy_that_reads_as_the_own_key_is_checked_as_a_copy, parametrized over the correct price and a wrong price of 45. sales.price is in neither the refused list nor the questions. Only the correct price is CONFIRMED.

How it was tested

• packages/flow: test_generate_coincidences*.py, test_generate_verify_contract*.py and test_zero_touch.py: 114 passed, against a local Postgres.
• Control: with the base demo_analysis.py, both new cases fail, because sales.price is refused as unchecked.
• mypy is clean on 591 sources. ruff check and format are clean on the changed Python files.

What couldn't be tested

• No live app or recorded demonstrations. The tests use synthetic SQLite and Postgres fixtures only.

[abrichr]

Independent adversarial review: APPROVE

From: Codex. Reviewed actual head 6d4a3a8 in a fresh detached worktree and locked environment.

packages/flow/openadapt_flow/generate/demo_analysis.py:709–716 classifies a demonstrated copy before considering the own-key coincidence. The resulting contract checks the source value. An actual own-key audit column still appears as a question and requires an explicit answer.

Validation: all 114 author-tested generator/coincidence/zero-touch tests pass with PostgreSQL 16. Full Flow Ruff/format and mypy pass (591 sources). Nine independent live generated-contract probes pass: correct price 40, five wrong prices, a changed source price of 55 versus stale 40, and the explicit own-key answer. The unchanged exact-parent fixture accepts all five wrong prices; the correct-price control passes there. This reproduces the original gap and confirms the new contract closes it.

Locked dependencies plus pinned psycopg2-binary 2.9.10; synthetic fixtures and private test state, headless resource gate. No live application or recorded demonstrations were tested. I did not edit or push this PR. CI must pass at the exact head before merge.