Add prompt dedup A/B simulation harness

[SecretSettler]

Summary

• Add offline-only system/skill prompt dedup A/B simulation to the Hermes opportunity analyzer
• Split candidate classes: same-type skill prompt, same-type system prompt, and cross-type system/skill
• Report simulated before/after char deltas and opt-in tokenizer deltas when an explicit backend is configured
• Keep all figures advisory/simulation-only; no runtime replacement, canonicalization, summarization, or prompt mutation
• Add CLI flags: --disable-prompt-dedup-ab and --prompt-dedup-tokenizer
• Document this as the evidence gate before canary replacement

Verification

• /usr/local/lib/hermes-agent/venv/bin/python -m pytest tests/test_hermes_context_opportunity_analyzer.py -q
• /usr/local/lib/hermes-agent/venv/bin/python -m pytest tests/test_hermes_plugin_patch.py tests/test_hermes_context_opportunity_analyzer.py tests/test_contextpilot_savings.py tests/test_contextpilot_savings_skill.py tests/test_hermes_contextpilot_monitor.py -q
• /usr/local/lib/hermes-agent/venv/bin/python -m pytest -q
• /usr/local/lib/hermes-agent/venv/bin/python scripts/analyze_hermes_context_opportunities.py --since-hours 24 --out-dir /tmp/contextpilot_prompt_ab_check --date prompt_ab_check --top-n 5
• Claude Code read-only review: Approve, no blockers

Stacked on PR #53, which is stacked on PR #52.