Keep admitted Nexus calls pending during cold replay

[rmcdaniel]

Fix

An admitted, completion-waiting Nexus call remains pending at its original yield. Cold replay now validates the recorded operation and submits no new start_service_operation command.

The published Python 2.6.0rc2 baseline in Server #425 appended 64, 63, 68 and 70 ServiceCallStarted events for one original call per cancellation case. Targets were started once. This fixes the replay readmission path responsible for the growing history.

Existing terminal delivery, async/acceptance-only calls, external results, changed-target rejection and legacy SideEffectRecorded replay retain their tests. Adds an immutable pending cancellation replay fixture.

Verification

• 70 focused Nexus and replay-corpus tests pass in Python 3.12.
• Ruff and mypy pass for the changed source.
• Three regressions fail against unchanged main and pass with the correction.
• Hosted regression-corpus and source checks must pass before merge.

This is one correction toward #116 and Workflow #753. It does not advertise Nexus cooperative cancellation or qualify a published replacement package. The accepted cancellation request regression intentionally requires no delivery until the separate capability-gated correction exists.

Related: durable-workflow/workflow#753