Repair inventing article sections instead of dropping them

[telivity-otaip]

Why

In the article channel a section that invented was regenerated once on a shorter budget and then dropped. The retry told the model to invent less without telling it what it had invented, and any drop set invent_reject for the whole article — so a stitch whose remaining text was clean was disqualified for a gap in coverage rather than for a fabrication.

On the article holdout after #50 that disqualified three items of four, including one arm that stitched zero words and was scored against a control that had written 135.

What changed

• Invent repair pass (section_repair_directives): the regenerate names the fabricated entities and figures the section has to remove, forbids replacements, and asks for a shorter section. Naming offenders is a calculated risk with this model, which is why the scrub stands behind it.
• Mechanical scrub (scrub_invented_sentences): if the repair still invents, the sentences carrying the offenders are cut and the section is re-checked against invent + parrot. The unit of removal is the sentence, not the span — cutting only the name out of "Fabrikam shipped the migration in nine weeks" leaves the fabricated claim standing with its subject missing. Lineation survives, so a scrubbed section keeps its paragraph rhythm.
• Drop is last, not second. A section is omitted only when repair and scrub both leave nothing usable.
• Final invent flag now fires when the stitch is empty or the stitched text invents against the visible brief — not because a section was dropped. Dropped-section offenders are still reported, under their own keys.
• Control arm parity. Both arms build sections through one draft_section_with_repair helper. A fact-lock only the product arm had to survive would separate the arms by editing policy rather than by writing.
• Receipts gained repair/scrub/drop counts and base_attempts, as counts only — the outline titles are mined from the holdout body.
• README: the article channel and eval paragraphs describe repair → scrub → drop.

Holdout result

Same carve, same model, k=2, real MLX run:

	post-#50	this branch
wins (article / base)	1 / 3	3 / 1
invent disqualifications	article 3 of 4	article 0 of 4, base 0 of 4
empty article drafts	1	0

Repair and scrub both fire on both arms, and the two items that previously lost only to an invent DQ are now the two best cadence distances in the run.

The verdict stays not_supported for one remaining reason: margin_within_chance. With four holdouts a 3–1 margin reaches p=0.3125, and even a clean sweep of four cannot clear alpha=0.10 by much — that is the carve size, not a fabrication problem, and the receipt says so rather than shipping a claim.

Test plan

• pytest -q — 377 passed
• ruff check src tests
• scripts/sanitize_check.sh
• Contoso coverage: repair keeps the section and names the offenders; scrub keeps the clean sentences when repair fails; a section is emptied only when both fail; a dropped section does not disqualify a clean stitch; the control arm repairs, scrubs, and still disqualifies when it cannot be repaired.
• Real PP_MLX_ALLOW=1 eval-write-article on the pinned carve; Contoso-safe receipt only.

Made with Cursor