Measure World correction revision stability

[eamirian]

What changed

• creates actual prior, treatment, and control BriefV1Alpha1 records over the exact FCC correction pair from P2C7
• replaces the affected instruction claim with the same exact correction claim in both revisions
• holds source coverage, claim count, prior/correction lineage, and reviewed workflow constant
• preserves both unaffected claim identities in treatment while the paraphrase-drift control preserves neither
• records exact stable, preserved, drifted, and unexpected claim sets in independently authenticated observed results
• classifies two matched pairs as useful and emits only a non-effective, non-selectable promote proposal
• adds duplicate-claim, incomplete-partition, missing-correction-visibility, invented-score, replay, and fresh-host controls
• records a bounded work packet, immutable audit evidence, and restrained candidate roadmap status

Why

ACE 0.6 needs revision-stability evidence in addition to citation correctness, contradiction attention, and detection delay. This packet proves that a product-owned rule can distinguish a bounded correction from gratuitous unrelated churn using unchanged Core + Intelligence measured-impact contracts.

This is stacked on World PR #10 and the Core measured-intelligence stack (#88, #89, #90). It does not merge, release, apply the proposal, pass SI4, close Core issue #38, or claim ACE 0.6 complete.

Verification

• focused P2C3-P2C8: 20 passed
• full World suite: 103 passed
• connector suite: 26 passed
• release contract: 7 passed
• public ace-core==0.5.0 boundary: 82 passed, 21 explicit candidate skips
• scoped Ruff check/format and git diff check: pass
• package build: unchanged 0.9.0 sdist and 45-entry inert data-only wheel; no Python or entry points; ace-core>=0.5.0,<0.6 unchanged

Repository-wide Ruff remains an inherited blocker: 15 lint findings and 20 pre-existing format targets, unchanged from the P2C7 parent.

Claim boundary

The paraphrase control is a frozen World fixture classification, not a general semantic-equivalence engine. This proves one recorded correction pair under one product rule; it does not prove live revision, population stability, calibration, source independence, legal truth, causality, general Brief quality, or human benefit. The proposal is not applied.

Core issue #49 F1, F3, and F5 still need explicit 0.6 release-owner disposition; this packet does not expand scope to implement or re-date them.

[eamirian]

Superseded by the direct-to-main convergence in #17 and the published World Intelligence 0.10.0 release. The release lineage composes the P2C3–P2C10 work and preserves its evidence; this stacked draft is now closed for repository hygiene.