fix(workflows): preserve pending result deliveries

[tt-a1i]

Summary

• retain every failed Workflow completion in memory before attempting durable delivery-state persistence
• drain an additional batch when a completion or lifecycle flush request arrives during an in-flight delivery
• isolate per-envelope persistence failures so one run cannot drop or block its siblings
• restore held-inline completions into the retry queue even when the pending-state write fails

The retry loop remains event-driven: a transport failure does not request its own immediate retry, so an unavailable host cannot create a busy loop.

Validation

• bun run check
• bun run test: 897 Node tests and 30 Vitest tests passed
• focused result-delivery regression suite: 8 passed
• two-axis review: Standards and Spec

Scope

This fixes the shared result-delivery queue defects in #107, #108, and #109. It does not attempt the separate multi-artifact transaction work in #110 or close the broader historical umbrella #71.

Closes #107
Closes #108
Closes #109