Preserve pending operations across signal wakeups

[rmcdaniel]

Problem

Addresses #96, delivered in 2.4.1 after published-artifact verification.

A signal waking a workflow with a pending scalar timer, remote activity or child
caused replay to emit that scheduling command again. Server allocated another
operation sequence. Timers acquired a later deadline, and the duplicate result
could collide with the next workflow operation during replay.

Change

• Validate the pending recorded operation and wait for its result without
  reissuing it. New operations and local activity preparation keep their paths.
• Include recorded timer delays in replay diagnostics, retaining histories
  without a recorded delay.
• Add a minimal append-only Avro replay fixture and focused regression coverage
  for repeated signals, fresh worker replay, original completion, changed code,
  legacy history and refusal to hide already duplicated timers.
• Update older test assertions that expected scalar rescheduling. Their
  cancellation identity, original authority, deadline and policy checks remain.
• Prepare patch 2.4.1 and document diagnosis of previously affected histories.

Verification

The unmodified published SDK reproduced duplicate timers on published Server
2.5.5 before image replacement and then failed replay on Server 2.5.6. The three
new scalar regression cases also fail against the unchanged source.

Local tests pass: 2,454 non-integration tests, plus 49 final regression/corpus
cases after adding the damaged-history refusal. The local corpus validator
passes with one new replay fixture. Ruff and mypy pass.

CI on c209731093ef8aff51e099ccd635ea5d6248eb33 passes Python 3.10, 3.11 and
3.12, regression corpus, lint, package build/smoke, the developer portal and
public boundary checks. Connected integration passes:
https://github.com/durable-workflow/sdk-python/actions/runs/37582960588

Review checks the changed-code diagnostics before omitting a recorded command,
retains local activity preparation and leaves cancellation delivery ahead of
ordinary waiting. Shielded cleanup timers validate the original receipt and
authority without starting another timer. Existing damaged histories continue
to fail explicitly rather than being silently repaired.

Published 2.4.1
from merged source 7cae7f7395aded953d9f950442798c8e450cd4fc. Main qualification
and protected publication pass. Installed-package checks against immutable
Server 2.5.5 and 2.5.6 preserve the original pending timer, history, run identity
and 204800-byte external payload through signals and fresh workers. The waiting
run completes with total eighteen and one timer schedule/fire. Completed and
queued controls also complete correctly. The broader image exercise requires
PHP/Rust worker restarts and a guarded observer recovery.

Delivery details.
Issue #96 is closed after verification, and the merged GitHub branch is deleted.