fix: prevent infinite auto-compact loop for unknown 3P models (#635)

[Vasanthdev2004]

Summary

Fixes #635 — auto-compact fires on every message for 3P models not in the context window table.

Root cause

The 8,000 token fallback in getContextWindowForModel() was too low. Combined with output token reservation (8k–20k) and the 13k auto-compact buffer, the effective context window went negative, causing isAboveAutoCompactThreshold to be true for any non-zero token usage.

Changes

1. Raise fallback from 8k → 128k (src/utils/context.ts)

• Unknown 3P models now get OPENAI_FALLBACK_CONTEXT_WINDOW = 128_000 instead of 8,000
• 128k is a reasonable conservative default for modern 3P models (most support 128k+)
• Effective context: 128k - 20k = 108k → healthy positive threshold

2. Add safety floor in getEffectiveContextWindowSize() (src/services/compact/autoCompact.ts)

• Math.max(effectiveContext, reservedTokensForSummary + 13_000)
• Even if someone sets CLAUDE_CODE_AUTO_COMPACT_WINDOW to a very low value, the effective context can never go below the summary reservation plus buffer
• This is a defense-in-depth guard: even if the fallback changes again, the floor prevents infinite loops

3. Add missing MiniMax model entries (src/utils/model/openaiContextWindows.ts)

• MiniMax-M2.7-highspeed, MiniMax-M2.5, MiniMax-M2.5-highspeed, MiniMax-M2.1, MiniMax-M2.1-highspeed (and lowercase variants)
• All at 204,800 context / 131,072 max output per MiniMax docs
• These were falling through to the 8k fallback because MiniMax-M2.5 doesn't prefix-match MiniMax-M2.7

4. Tests

• Updated existing test: unknown 3P model → 128k (was 8k)
• Added test: MiniMax M2.5/M2.1 variants resolve to 204,800 context
• Added autoCompact.test.ts: getEffectiveContextWindowSize never returns negative, getAutoCompactThreshold never returns negative

Validation

+5 files, +100/-9

CI should pass — all changes are additive or value corrections, no breaking API changes.