fix: reset metrics registry per benchmark cycle

[xinlian12]

Problem

When running the benchmark with -cycles 2+, the Cosmos SDK calls registry.clear() + registry.close() via ClientTelemetryMetrics.remove() when a CosmosClient is destroyed between cycles. Since the benchmark passed the same CompositeMeterRegistry to every cycle's client, closing cycle 1's client wiped all meters from the shared registry. Cycle 2's CosmosClient created new meters (with a new ClientCorrelationId), but the registry was in a broken state — no CSV data was written for cycle 2.

Root Cause

ClientTelemetryMetrics.remove(registry) calls:

1. registry.clear() — removes all meter instances from the registry
2. registry.close() — shuts down the registry so it stops accepting new meters

Both propagate through the CompositeMeterRegistry to all children, destroying the SimpleMeterRegistry that CsvMetricsReporter reads from.

Fix

Create a fresh, disposable metrics pipeline per cycle:

1. Each cycle creates its own CompositeMeterRegistry + SimpleMeterRegistry + CsvMetricsReporter
2. The LoggingMeterRegistry (console output) is added to each cycle's registry but disconnected before shutdown so the SDK's clear()/close() doesn't destroy it
3. Reporters are flushed before client shutdown to capture all data
4. The SDK can safely destroy the disposable cycle registry — nothing shared is affected
5. Each cycle's CSV files start with fresh counters (count=0)

Verified

Tested with -cycles 2 -settleTimeMs 30000 on 1t-c16-ReadThroughput-http1:

• Before fix: Only ClientCorrelationId.00001 CSV files existed (cycle 1 data only)
• After fix: Both 00001 and 00002 CSV files exist with complete, independent data

	Cycle 1 (00001)	Cycle 2 (00002)
CSV rows	20	20
Total ops	1,531,107	1,290,714
Mean latency	1.85ms	2.20ms
Starts fresh	count=0	count=0

[Copilot]

Pull request overview

Fixes multi-cycle benchmark runs in azure-cosmos-benchmark where closing a cycle’s CosmosClient would clear()/close() the shared Micrometer registry (via ClientTelemetryMetrics.remove(...)), preventing subsequent cycles from emitting CSV metrics.

Changes:

• Moves metrics registry + destination reporter creation into the per-cycle lifecycle loop to ensure each cycle uses a fresh, disposable registry.
• Keeps console logging (LoggingMeterRegistry) shared across cycles, and detaches it before client shutdown to avoid being cleared/closed by the SDK.
• Stops (flushes) CSV/CosmosDB reporters before client shutdown so final interval data is written before the registry is destroyed.

[xinlian12]

The failed tests: c.a.c.CosmosContainerChangeFeedTest.CosmosContainerChangeFeedTest -> does not related to changes in this PR

[xinlian12]

/check-enforcer override