Fixes NousResearch#40803: Infinite Context Compaction Loop (messages=N->N) on low context_length / limit configurations

When summary_target_ratio is set high (e.g., 0.45) with a constrained context_length,
the tail budget becomes so large that the compression window shrinks to <5 messages.
Compressing only 1-2 messages yields 0 token savings (due to summary overhead),
triggering compression again on the next turn, creating an infinite loop.

This fix detects when the compression window is too small (<5 messages) and would
result in minimal token savings (<2000 tokens). When detected, it falls back to a
more aggressive tail budget (50% of original) to create a larger compression window.

New safety checks prevent:
1. No-op compressions that save 0 tokens
2. Infinite compression loops on constrained configs
3. Wasted API calls and latency on every message turn

Added 6 comprehensive tests verifying:
1. Small windows trigger fallback
2. Fallback creates larger compression window
3. Large windows don't unnecessarily trigger fallback
4. Fallback prevents no-op compression
5. Window size calculation is correct
6. Token estimation in windows is accurate

All 6 tests pass.

Fixes NousResearch#40803