This release separates automated regression checks from observed native behavior.
- 257 local checks: engine/DOM 105, favicon 19, extension 36, relay 48, MCP tools 30, OAuth 8, legacy Web 11. Both Worker TypeScript checks pass. Node24 and the supplied lockfiles are used.
- Native core on 0.2.2: 11/11: password redaction/edit refusal, exact Enter output, completed scroll gesture, HTML drag/drop, navigation, stale reference refusal, history, bounded network metadata and local confirmation without submission.
- Native background on 0.2.3: 8/8: inactive new tab despite legacy active:true, unchanged active public tab before/after creation and input, exact form result, controlled target marker, original title retained and unrelated tabs not claimed by listing.
- Native close on 0.2.3: two disposable fixture tabs returned closed:true after Chrome's actual remove ACK; both were absent from the following inventory.
- ChatGPT on 0.2.2: a real connected conversation filled a field, selected Green, clicked and inspected a returned image. Text and screenshot pixels were independently checked.
- Stop/timeout regression: the full engine runs against a simulated remove call that never resolves. Stop and timeout release the extension queue, allow the next command and never retry remove. The old unbounded implementation fails this regression.
The input paths are unchanged between 0.2.2 and 0.2.3; 0.2.3 changes close acknowledgement and cancellation. It does not imply every native core case was rerun through ChatGPT on 0.2.3.
One initial 0.2.3 background measurement saw an empty active-public-tab inventory become a public active tab during the run; two comparison checks failed. The complete subsequent run passed 8/8 with a public active tab present both before and after. The first transition cannot establish its cause and is not counted as a pass.
On five verified form runs, 0.2.1 median times were 1098 ms for individual MCP calls and 1044 ms for a batch. A comparable Codex high-level API path had a 1009 ms median. Model thinking is excluded. Different APIs and a small sample do not establish universal speed or capability parity.
No desktop foreground-focus observation, mobile E2E validation or continuous YouTube audio/video claim is made. The public demo capture contains only a disposable fixture, with no personal page data. Account IDs, deployed URLs, pairing tokens and private conversation links are omitted. Dependency installation currently reports npm advisories in the pinned Worker dependencies (7 high for Web Bridge and 3 high for the YouTube Worker). This publication preserves the tested lockfiles; it is not a dependency-security audit or a claim that the advisories were remediated.