feat: auto-detect 1M context window and scale guard thresholds

[ynaamane]

Summary

Claude Code now appends [1m] to model IDs when using extended 1M context (e.g. claude-opus-4-6[1m]). Cozempic v1.0.0 hardcodes all models to 200K in MODEL_CONTEXT_WINDOWS, causing:

• Premature guard pruning: hard threshold fires at 150K tokens (75% of 200K) instead of 750K (75% of 1M)
• Incorrect diagnostics: cozempic current --diagnose shows 50% context used when it's actually 10%
• Aggressive overflow recovery: danger threshold of 90MB is too low for 1M sessions that can legitimately reach 200-400MB

Changes

• tokens.py: Add [1m]-suffixed entries to MODEL_CONTEXT_WINDOWS (1M context). Update detect_context_window() with bracket-aware prefix matching for versioned model IDs (e.g. claude-opus-4-6-20260301[1m])
• init.py: Update SessionStart hook template to extract context_window.context_window_size from Claude Code stdin JSON and pass --context-window dynamically to guard daemon
• guard.py: Always detect context window upfront (used for display + overflow scaling). Display detected context window in guard banner. Scale overflow danger_threshold_mb and danger_threshold_tokens based on context window
• tests/test_model_detection.py: Add 15 new tests covering 1M exact match, versioned prefix match, 200K/1M non-confusion, env override priority, threshold scaling, and context % calculations

Backward compatibility

• Zero breaking changes: 200K models keep their exact same thresholds and behavior
• COZEMPIC_CONTEXT_WINDOW env var override still takes highest priority
• Old hooks (without --context-window) fall back to model-based detection from session JSONL
• Users with existing hooks can re-run cozempic init to get the updated template

Before/After

Metric	Before (200K assumed)	After (1M detected)
Hard token threshold	150,000	750,000
Soft token threshold	90,000	450,000
Context % for 100K tokens	50%	10%
Overflow danger (MB)	90	90 (scaled to ~1.8x hard MB threshold)
Overflow danger (tokens)	not set	900,000 (90% of context)

Test plan

• All 89 existing tests pass
• 15 new tests pass (1M detection, threshold scaling, edge cases)
• Manual: cozempic current --diagnose with 1M model shows correct context window
• Manual: Guard banner shows Context: 1.0M on 1M sessions
• Manual: Hook extraction test: echo '{"session_id":"test","context_window":{"context_window_size":1000000}}' | bash -c '<hook_command>'