Summary

Salvage of PR #4156 by @SHL0MS (cherry-picked onto current main) with one test fix.

Three targeted improvements to context compression quality, addressing the most impactful issues from #499.

1. Higher serializer truncation limits (3K → 6K)

The summarizer model was working from heavily truncated data. Magic numbers replaced with named class constants:

These only affect input to the auxiliary summarizer model — no impact on the main model's context or prompt cache.

2. "Tools & Patterns" section in compression prompts

Both prompt templates (first-pass and iterative-update) now include:

## Tools & Patterns
[Which tools were used, how they were used effectively, and any
tool-specific discoveries. Accumulate across compactions.]

Preserves HOW tools were used effectively across compaction boundaries — the agent often re-discovers effective flag combos after compression.

3. Degradation warning after repeated compressions

After 2+ compressions:

⚠️  Session compressed 3 times — accuracy may degrade. Consider /new to start fresh.

compression_count was tracked internally but never surfaced to the user. Uses _vprint(force=True) — purely user-facing, does NOT inject into the message stream or affect prompt caching.

Follow-up fix

Set compression_count = 1 on the MagicMock in test_flag_reset_on_compression — the new degradation warning reads this attribute as an int, but the existing mock returned a MagicMock object.

Test plan

• 36 compressor tests pass
• 23 context pressure tests pass (including the fixed one)
• py_compile clean on both files

Ref #499