[Blind-Striker]

Canceling a pending terminal read previously passed cancellation to the physical WebSocket and could abort the connection. Both terminal families now keep receiving independently, so a consumer can cancel and later read or write on the same healthy attachment. Queued output precedes the terminal outcome; explicit disposal abandons unread output and joins owned work.

Persistent input preserves locally requested viewport dimensions in send order. Both connection options add a finite SendTimeout (30 seconds by default), covering queue wait and physical send. Queued expiry affects that call; physical interruption ends the attachment without retry. Runtime/packaging canon and ADR-0023 document these contracts.

The first hosted Windows run exposed nested cancellation scheduling in existing net472 cleanup test helpers. The follow-up uses the already-owned worker directly and removes unrelated observation timers from teardown unit tests. An isolated actual-code probe changed from stuck cleanup to one completed callback; all 31 focused ownership tests and the full 5,134-test local gate passed.

Validation:

• TDD red/green evidence for viewport ordering, send deadlines, pending-read reuse, stable outcomes, and shared disposal.
• Windows Release solution build, Slopwatch, whitespace/style checks, and full suite: 5,134 passed, 0 failed, 0 skipped.
• Linux real normal/persistent PTY tests on net8/net9/net10: 6 passed, 0 failed, 0 skipped; persistent daemon round trip and same-connection execution after normal PTY read cancellation verified.
• Before/after allocation benchmarks: 14 matched cases across net10/net472. For 1,024 small frames on .NET 10, normal PTY rose from 205,600 to 240,464 B and persistent PTY from 197,624 to 232,288 B per connection lifetime. This is the measured cost of the new contract; no speed claim.

The consumer queue is unbounded and can grow for slow or absent consumers. Final three-platform CI passed on b231755: Windows 5,134, Linux 4,030, macOS 4,030; zero failures/skips. All 20 normal/persistent terminal live outcomes across ten platform/framework legs were verified from the original TRX reports. The separate upstream observer-input question is outside this change.

Final documentation follow-up 62970f8 corrects the receive-buffer lifetime in README and roadmap. Its only changes since tested b231755 are those two Markdown files; CI is skipped for that follow-up under the repository workflow rule.