ConnectOnion 1.9.3rc17 (release candidate)

This document describes the candidate; availability requires its matching public PyPI package and GitHub prerelease. During preparation the public/installed baseline was 1.9.3rc16. Stable remains 1.9.2. Python 3.11+ and rc16's dependency floors are retained; no runtime dependency is upgraded by this candidate.

An explicit correction can reach a page without new raw material

Direct investigation can process fresh, eligible attributed feedback in one bounded maintenance turn when no new original-source material is available. The result distinguishes feedback_only, zero raw items and feedback inputs; it does not claim fresh research. The original author/basis and source IDs stay attributed. Refused, deferred, oversized or unaffordable work remains pending. Acceptance receipts belong to targets actually accepted; a no-change result must still match the selected page snapshot. A healthy empty source window is not an unavailable provider. This does not implement the broader external-agent, mixed-source or multi-round product proposals.

Old words need their original context to inherit an exemption

Retained prose is compared with its complete paragraph/bullet context, section hierarchy, meaningful word spacing, relative cited-source identities and order. A removed qualification, moved evidence or changed current/history placement does not qualify solely because some words remain. Source rows, headings and examples are not evidence that a claim is old. Deferred old-content findings remain insufficient, with their count, rather than acquiring a factual PASS. This is deterministic continuity, not semantic equivalence or proof of truth.

Recovery hints keep the argument names that were supplied

A tool error now reports its supplied argument names faithfully. This corrects the recovery hint without changing the tool's signature or approval policy.

Synthetic failure tests record their requested delays

Three recovery cases keep exact retry delays/final-attempt checks and retained partial/healthy-source assertions while replacing real sleeps in those tests. The same three-test command measured45.76s before and0.68s after. Full offline regression on the test-change implementation passed17,093tests; whole-CI savings are not established. No production retry delay, timeout, coverage floor or release gate is changed.

Acceptance and limits

The final version candidate still needs full offline, dependency/security, package/installed-wheel, independent review, version-PR CI and immutable-tag publication gates. See the scope and manual trial record. The reviewed backend visual exemption must describe changed data/diagnostics and unchanged reader structure; a passing manifest parser is not native proof. No current REM desktop/phone/source-popup/privacy/focus/accessibility screenshots, real init/provider/model/mail/scheduler trial or real notebook rewrite is claimed.

Pending reader/accounting/translation/navigation candidates retain their own acceptance gates and are not included. Processing receipts and tests do not establish source truth, useful first-run insights, overnight recall or product maturity. Existing saved pages are not automatically repaired by an installation.