fix(compressor): soften summary prompt for content filters

[LeonSGP43]

Summary

• Reword the context compressor summarizer prompt to avoid injection/handoff phrasing that can trip Azure/OpenAI-compatible content filters.
• Keep the existing structured checkpoint schema and safety constraints intact.
• Add a regression test that inspects the actual prompt sent to the summary LLM.

Why

Issue #19362 reports HTTP 400 content-filter failures during context compression with Azure/OpenAI-compatible providers. The previous prompt used phrases like "injected", "DIFFERENT assistant", and "Do NOT respond", which can resemble prompt-injection/jailbreak handling text even though the operation is internal summarization.

Verification

• scripts/run_tests.sh tests/agent/test_context_compressor.py - 67 passed, 4 existing event-loop deprecation warnings

Fixes #19362

[alt-glitch]

Likely duplicate of #19370 and #19381 — all three PRs fix the same content-filter false-positive in context_compressor.py (issue #19362).

[alt-glitch]

Likely duplicate of #19370 and #19381 — all three PRs fix the same content-filter false-positive in context_compressor.py (issue #19362).