fix: preserve legacy merge progress when a worker cannot connect

[beinan]

When an ordinary worker drains during deployment, its connection failure currently stops legacy WAL merge fan-out before later healthy workers and drops reclaimed-generation metrics for already completed workers.

Continue the serial fan-out only after a connection-establishment failure, retain completed progress, and leave the canonical task failed/retryable for unavailable shards. Stop on HTTP errors, lost responses and post-send timeouts because an unfenced legacy writer may still be running. The internal worker client disables redirects so a failed redirected POST cannot be mistaken for a request that was never admitted; only connection establishment has a two-second deadline.

Validation:

• Four real TCP regressions pass: healthy progress on both sides of a refused connection, HTTP failure stopping before the next writer, redirect to an unreachable endpoint even with a valid merge payload, and timeout after server admission.
• Full master unit suite: 70 passed, 84 existing isolated-etcd tests ignored, zero failed.
• Separately ran all eight merge-execution recovery tests against a fresh isolated local etcd: all passed, including orphan fencing, surviving-writer protection, partial success/retry exhaustion and no legacy downgrade.
• Formatting and diff checks pass. Remaining CI is running; merge only after every workflow passes.