fix: derive compaction threshold from context window for large (1M) models

[akhil29]

What

Fixes early-compaction on large-context models (e.g. Claude Opus 4.6/4.7/4.8 with 1M-token windows) and adds extensive bug-catching tests for the Vertex AI Anthropic provider.

The bug

On a 1M-token Opus model with default config, compaction fired at ~100K tokens — roughly 10% of the window — because the default token_threshold (100K) was applied as an absolute cap:

min(token_threshold.unwrap_or(100_000), 0.70 * context_window)
= min(100_000, 700_000) = 100_000

The fix (two parts)

1. forge_domain/src/agent.rs — compaction_threshold()

   • When token_threshold is unset (the realistic default), derive the threshold purely from the context window (70%), so large windows aren't capped to 100K.
   • When explicitly configured, it is still honored as an absolute cap: min(configured, 70% * window).

2. forge_app/src/dto/anthropic/response.rs — get_context_length()

   • The generic claude-opus-4- prefix wrongly captured the 1M-token claude-opus-4-6/4-7/4-8 models at 200K. Added an explicit 1M branch before the generic 200K branch. Older claude-opus-4 / claude-opus-4-1 remain at 200K.

Tests (genuine bug-catchers — red without the fix)

Extensive Vertex AI Opus compaction-threshold coverage across large (1M) and small (200K) windows:

• Derived-threshold assertions (1M → 700K, 200K → 140K)
• Below / at / above trigger boundaries (threshold - 1 does not trigger, threshold does)
• Cross-window guarantee: a 200K-sized context does not compact on a 1M window
• get_context_length coverage for the Opus 1M models
• Updated the stale threshold test that encoded the buggy capping behavior

Verified these fail on the unfixed logic and pass with the fix applied.

Verification

• forge_domain vertex_opus: 7 passed
• forge_domain compaction_threshold: 6 passed
• forge_app get_context_length: 8 passed
• Full forge_domain and forge_app suites: green

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

Akhil Appana seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}