fix(agent): rephrase compressor summariser preamble to avoid Azure content filter

[konsisumer]

Rephrase the context-compression summariser preamble so Azure/OpenAI-compatible content filters stop classifying it as a jailbreak attempt.

What changed and why

• agent/context_compressor.py: replaced the _summarizer_preamble (and the first-compaction prompt body that re-echoed the same framing) so it no longer contains the phrases the filters were triggering on:
  • "Your output will be injected as reference material for a DIFFERENT assistant"
  • "Do NOT respond to any questions or requests in the conversation"
  • "different assistant that will continue this conversation"
    These are negated meta-instructions telling the model to ignore the user — i.e., textbook prompt-injection / jailbreak shape — which produced deterministic 400 ResponsibleAIPolicyViolation (jailbreak: detected) on every auto-compress call against Azure/OpenAI-compatible deployments. The earlier [SYSTEM: → [IMPORTANT: mitigation ([Bug]: Azure OpenAI security filters block some Hermes requests due to trigger word #6576 / fix(prompts): rename [SYSTEM: → [IMPORTANT: in all user-injected markers (Azure content filter) #16114) handled bracketed markers; this is the remaining, distinct trigger flagged in the issue.
• New preamble frames the task neutrally as a summarisation job using positive imperatives ("Produce a structured summary document", "Begin your output directly with the first section heading", "Replace … with [REDACTED]"). Same intent, no jailbreak-shaped phrasing. Drops the "different assistant" handoff framing entirely — the next session is an implementation detail the summariser doesn't need to know about.
• Updated the module docstring so the design-rationale bullet points reflect the new wording instead of crediting OpenCode/Codex phrases that have been removed.
• tests/agent/test_context_compressor.py: new TestSummarizerPreambleContentFilterSafety class captures the prompt sent through call_llm for both the first-compaction and iterative-update paths and asserts none of the historical trigger phrases reappear, so future edits can't silently regress this Azure compatibility gap.

How to test

• pytest tests/agent/test_context_compressor.py -q (68 passed locally)
• pytest tests/agent/test_compress_focus.py tests/agent/test_compressor_image_tokens.py tests/agent/test_context_engine.py -q (35 passed)
• pytest tests/test_cli_manual_compress.py -q (1 passed)
• Manual: trigger an auto-compress against an Azure/OpenAI-compatible endpoint that previously returned content_filter / ResponsibleAIPolicyViolation on the auxiliary summary call; the call should now succeed instead of entering the 60-second summary-failure cooldown.

What platforms tested on

• macOS on darwin-arm64 (local)

Fixes #19362

[alt-glitch]

Likely duplicate of #19370 — same fix for Azure content-filter false-positive on context_compressor.py summarizer preamble. #19370 is already open with the same approach.

[alt-glitch]

Likely duplicate of #19370 — same fix for Azure content-filter false-positive on context_compressor.py summarizer preamble.

[konsisumer]

Closing — deferring to #19370 by @liuhao1024 which addresses the same. Reopen if that PR stalls.