feat: dynamic context pressure warning with model-aware thresholds

[tuanhung303]

Summary

• Add dynamic context pressure warning based on 70% of the agent's actual context window
• Show 🟢/🟡/🟠/🔴 {percent}% in status bar with | separator
• Reasoning block notifications show 🧠 (xN) instead of 🧠 "—"
• Periodic checkpoint reminder shows model info, remaining tokens, and savings from pruning

Changes

Core

• lib/strategies/utils.ts — MODEL_CONTEXT_WINDOWS map (Claude 200K, GPT 128K, Gemini 1M, DeepSeek 128K), resolveModelContextWindow(), getEffectiveContextThreshold(), getContextStatus()
• lib/state/types.ts — ContextPressure interface on SessionState (transient, recomputed each turn)
• lib/hooks.ts — Compute context pressure after syncSessionState, populate state.contextPressure

Config

• lib/config/schema.ts — Replace maxContextTokens with fallbackContextWindow (200K) + warningThresholdPercent (0.7)
• lib/config/defaults.ts — Updated defaults

UI

• lib/ui/utils.ts — formatStatsHeader() accepts SessionState, appends status emoji + percent, uses | separator
• lib/ui/notification.ts — Pass state to formatStatsHeader()
• lib/ui/minimal-notifications.ts — Reasoning blocks show 🧠 (xN) not 🧠 "—"
• lib/ui/pruning-status.ts — Same reasoning fix in formatItemizedDetails()
• lib/messages/todo-reminder.ts — Rich checkpoint: model info, status label, remaining tokens, savings

Tests

• Updated 8 test files with new contextPressure state and config fields
• All 557 tests pass

Example Output

Status bar:

「 💬 12(6K) ▼ | 🧠 141(282K) ▼ | ⚙️ 177(143.8K) ▼ | 🟡 59% 」- 🧠 (x4)

Checkpoint reminder:

⚡ Context: 59% 🟡 Elevated — 82.0K tokens remaining
📋 118.0K / 200.0K (200.0K context window)
🤖 Model: Claude Opus (200.0K context)
💾 Savings: 12.3K tokens reclaimed via pruning