Publish the namespace isolation experiment

[rmcdaniel]

Summary

• add a reusable, published-artifact-only noisy/control namespace experiment
• exercise API, workflow/activity dispatch, timers, signals, queries, updates, schedules, child workflows, Nexus, replay-heavy history, and external payload pressure
• disrupt Redis and Server during load, retain queue/operator/resource evidence, and require post-pressure recovery
• document the measured shared-cell envelope and the five-producer boundary where application-level admission no longer protects control latency

Recorded runs

Default cell, Server 2.0.3 and Python SDK 2.0.0:

• two independent passing runs
• representative retained run: 7 uncontaminated control completions, p50 10.1554s, p95/max 11.2527s
• 10 concurrency, 8 request-rate, and 2 open-run noisy-namespace rejections
• Redis recovery 4.532s; Server recovery 15.095s
• control/noisy post-pressure completion 2.878s / 2.7909s
• peak sampled stack memory 706.888 MiB; Server memory 157.5 MiB; aggregate CPU 185.6%

Boundary run with five continuously retrying noisy producers:

• failed the same cell as intended
• one uncontaminated control completion at 16.9974s
• proves PHP-level admission is not volumetric DoS protection

Local validation

• bash -n scripts/perf/run-namespace-isolation.sh
• python3 -m py_compile scripts/perf/namespace_isolation.py
• ruff check scripts/perf/namespace_isolation.py
• git diff --check

Refs #91