Add configurable context window limit / auto-compaction threshold

Problem

Users on plans that default to 1M context windows have no way to opt into a smaller effective context window (e.g., 200K). This means auto-compaction only triggers near the 1M limit, which:

• Burns through token quotas faster — every request sends up to 1M tokens of context
• Slows down responses — more context = more processing time
• Reduces total work per session — tokens are spent on stale context instead of new work

For power users doing long iterative sessions, this is a significant cost/efficiency issue.

Proposed Solution

Add a configurable context window limit, e.g.:

Option A: settings.json

{
  "contextWindowLimit": 200000
}

Option B: CLI flag

claude --context-window 200000

Option C: Both

This would cap the effective context window and trigger auto-compaction at the configured limit instead of the model maximum.

Use Case

• User has Opus 4.6 with 1M context by default
• User wants to compact aggressively at ~200K to maximize productive work per token budget
• Currently the only workaround is manually running /compact, which is easy to forget

Alternatives Considered

• Manual /compact — works but requires user discipline and is easy to forget
• Using subagents — helps but does not solve the core issue
• CLAUDE.md instructions to the model — unreliable since the model cannot actually measure its own context usage precisely