bench: eliminate intermittent EADDRNOTAVAIL in repeated compare runs

[crrow]

Problem

Repeated redis-bench compare runs intermittently fail with:

• dial tcp 127.0.0.1:6390: connect: can't assign requested address
• dial tcp 127.0.0.1:6391: connect: can't assign requested address

This makes stability validation noisy and weakens CI reproducibility.

Suspected Cause

Current benchmark path opens a very high number of short-lived TCP connections in tight loops, causing local port/socket pressure.

Scope

• Profile connection lifecycle in benchmark workers
• Reuse persistent connections per worker (or bounded connection pool)
• Ensure clean server process teardown between runs
• Add repeat-run validation command/path

Acceptance Criteria

• Run redis-bench compare --requests 2000 --concurrency 30 5x sequentially without EADDRNOTAVAIL
• Keep benchmark report generation unchanged
• just check and related tests pass

Priority

• P1 (reproducibility)
• token budget: M

[crrow]

Leader dispatch:

Please focus on deterministic repeated-run behavior.

Required execution plan:

1. Add a micro benchmark harness or test helper to run compare repeatedly.
2. Refactor benchmark request execution to reuse worker connections.
3. Verify no EADDRNOTAVAIL across 5 sequential runs at 2000/30.
4. Keep report schema/outputs backward compatible.

Validation required in PR:

• just check
• repeat-run evidence (5 successful runs)