fix: add Kimi K3 1M context window to DEFAULT_CONTEXT_LENGTHS

[datachainsystems]

Summary

Kimi K3 ships with a 1M-token context window (verified: platform.kimi.ai/docs/overview) but was falling through to the generic "kimi": 262144 catch-all in DEFAULT_CONTEXT_LENGTHS.

Change

• Added "kimi-k3": 1_000_000 before the "kimi": 262144 catch-all in agent/model_metadata.py
• Longest-key-first substring matching ensures K3 resolves to 1M while older/unknown Kimi models still hit the 256K default
• Added test_kimi_k3_context_1m covering native, vendor-prefixed (kimi/, moonshotai/), and older model (kimi-k2.6, kimi-k2) fallback verification

Verification

tests/agent/test_model_metadata.py::TestDefaultContextLengths::test_kimi_k3_context_1m PASSED
tests/agent/test_model_metadata.py - 122/122 passed, 0 failed

Pattern

Same longest-key-first approach used by DeepSeek V4 (#...), GLM-5.2 (#...), and others.

[alt-glitch]

This was generated by AI during triage.

Related to open #67115. Both address Kimi K3 context sizing, but this PR changes the global default fallback while #67115 keeps the repair scoped to canonical Kimi Coding endpoints; please choose the intended contract.

[datachainsystems]

Thanks for the triage. I reviewed #67115 and believe #67685 is the correct approach — here's why:

#67115 fixes _endpoint_scoped_context_length which only triggers for api.kimi.com/coding endpoints. But the default base_url in the kimi-coding plugin is https://api.moonshot.ai/v1 — where most users access Kimi K3. That path would still fall through to 262K.

#67685 fixes DEFAULT_CONTEXT_LENGTHS globally — the same mechanism used by DeepSeek V4, GLM-5.2, Qwen3.6, and every other model family. The model's context window is a property of the model, not the provider. Kimi K3 has 1M context regardless of which endpoint you reach it through.

The endpoint-scoped path already handles the bare k3 slug specifically for api.kimi.com/coding. The DEFAULT_CONTEXT_LENGTHS entry handles the public-facing kimi-k3 slug for all endpoints. These are complementary, not competing — both should land.

If there's a concern about legacy Moonshot keys that don't serve K3 returning 1M for kimi-k3: those keys wouldn't expose a kimi-k3 model in the first place, so the fallback is never reached for them.

[teknium1]

Merged via PR #68108 — your commit was cherry-picked onto current main with your authorship preserved in git log. Thanks for the fix! The salvage also widened the rollout to the remaining Kimi-direct catalog surfaces (curated picker lists, setup defaults, and bare-slug family classification).