feat: model-aware safety buffer for context compaction (closes #2)

[Blankll]

What

Implements #2 — replaces the flat compaction safety buffer and ratio with tiered, model-aware functions.

Changes

Safety buffer

• Removed flat SAFETY_BUFFER_TOKENS = 13,000
• Added safety_buffer_for_window(context_window) -> usize:
  • ≥ 200K → 8,000
  • ≥ 32K → 4,000
  • < 32K → 2,000

Compaction ratio

• Removed flat DEFAULT_COMPACT_RATIO = 0.75
• Added compact_ratio_for_window(context_window) -> f64:
  • ≥ 100K → 0.85 (less aggressive summarization)
  • < 100K → 0.75

API change

• compact_trigger_threshold(&ModelSpec) — now model-aware, computes usable_window() internally

Impact

• Large windows (GPT-4.1 1M+) use more capacity before compacting
• Small windows (Ollama 8K) no longer trigger at near-zero tokens (was 1 → now 4,144)
• No downstream breakage: crate not yet published, all callers updated

Verification

• ✅ Build: green
• ✅ Tests: 62/62 pass (8 new)
• ✅ Clippy + fmt: clean
• ✅ Oracle peer-reviewed