fix(xai): drop stale 256K grok-4.6 context cache

[Julientalbot]

Summary

Official card (docs.x.ai/developers/models/grok-4.6): 500,000 context.

#84661 landed the catalog. main already lists native grok-4.6 on the xAI picker. This PR is only the leftover cache guard.

Pre-catalog builds resolved grok-4.6 via the grok-4 catch-all (256K) and persisted it. After the catalog land the table is 500K, but a leftover cache entry still wins. Same pattern as the existing grok-4.3 guard.

Not in this PR

• No catalog restatement
• No picker / extras restatement (already on main)
• No xhigh special-case
• No tool-default flip 4.5 → 4.6

Test plan

• uv run --extra dev pytest tests/agent/test_model_metadata.py::TestGrok46StaleCacheGuard tests/agent/test_model_metadata.py::TestGrok43StaleCacheGuard tests/agent/test_reasoning_stale_timeout_floor.py -q → 40 passed
• CI green

[lihongjie224]

Following this — grok-4.6 is already in the live xAI / models.dev catalog, but without this PR Hermes still treats it as the grok-4 catch-all (256K context, no reasoning.effort, no stale-think floor).

Any estimate on when this is likely to merge?

[Julientalbot]

Landed on main via #84661 (3e09adb, merged 2026-08-12 17:37Z) — same catalog surface this PR was opened for:

• DEFAULT_CONTEXT_LENGTHS["grok-4.6"] = 500000
• _GROK_EFFORT_CAPABLE_PREFIXES includes grok-4.6
• reasoning stale-timeout floor 300s

This branch is now CONFLICTING against that merge. Closing as superseded rather than rebasing a duplicate.

@lihongjie224 no remaining wait on this thread: pull current main (or until then model.context_length: 500000).

Delta this branch still had that #84661 did not: invalidate a persisted 256K cache leftover for grok-4.6 (same guard as grok-4.3). Only worth a tiny follow-up if that still bites after the catalog land.

[Julientalbot]

Reopened. Closing this against #84661 was a mistake on my side.

#84661 landed the catalog (500K / effort / floor) with no description and flipped the OpenRouter picker x-ai/grok-4.5 → x-ai/grok-4.6. It did not cover:

• stale 256K cache invalidate (this branch, same pattern as grok-4.3)
• native grok-4.6 on the xAI picker fallback/extras
• catalog + cache tests

Rebuilt onto current main (e0f6c65e1). No longer a conflicting duplicate.

[Julientalbot]

Rebuilt onto current main (9deb0302c).

Picker leftover is gone — main already has native grok-4.6 in _XAI_STATIC_FALLBACK / extras. This branch is now cache guard only (clone of grok-4.3) + TestGrok46StaleCacheGuard.

[teknium1]

Merged via PR #85434 — your commit was cherry-picked onto current main with your authorship preserved in git log (merged 91e550b). Thanks for self-pruning the parts #84661 superseded and keeping this a clean cache-guard delta, and for the E2E test.

On top of your commit we generalized the guard: the per-model _model_name_suggests_* predicates (grok-4.3, grok-4.6, minimax-m3) are now one generic _stale_pre_catalog_cache_entry() driven by a _PRE_CATALOG_STALE_KEYS set, so future pre-catalog leftovers only need a one-line catalog entry. Your TestGrok46StaleCacheGuard E2E test is preserved as-is.