Fix context window overflow: auto-recovery and proactive compaction

[yunus25jmi1]

Problem

When tool results push the conversation past the model's context limit, the agent crashes with raw HTTP errors like \prompt is too long: 226360 tokens > 200000 maximum\ and never recovers.

Fixes #1750

What this does

Prevents overflow before it happens:

• After tool calls return, estimates how many tokens the new messages add
• If the total would exceed 90% of the context limit, compacts automatically before the next API call

Recovers if overflow still occurs:

• Detects overflow errors from all major providers (Anthropic, OpenAI, Google, Bedrock — 14 patterns)
• Stops wasting retries on errors that won't go away
• Auto-compacts the session and retries with the shorter context
• Shows a clear message instead of raw HTTP/JSON errors

Files changed

• \pkg/runtime/fallback.go\ — overflow detection, skip retries, error wrapping
• \pkg/runtime/runtime.go\ — auto-recovery loop, proactive compaction, token estimation, friendly errors
• \pkg/runtime/fallback_test.go\ — 26 new test cases
• \pkg/runtime/runtime_test.go\ — 7 new test cases

Testing

• All existing tests pass
• 33 new test cases added
• Linter: 0 issues