Code Review — fix(installer): don't flag SD/embedding models as 'Context unverified'

Summary

Solid, well-scoped bug fix — the root cause is correctly identified and the fix is minimal and clean. The misleading ⚠️ Context unverified! on the first model of an SD profile came from __init__ seeding self._ctx_verified = None, which made the first loop iteration's hasattr(self, "_ctx_verified") check pass even for SD/embedding models that never run verification. Moving the reset to the top of each model iteration (and dropping the seed from __init__) makes the warning fire only when an LLM with a min_context_size was actually checked and the server reported no context back. I ran the 3 new tests locally — all pass.

The one thing worth knowing: I confirmed every read of _ctx_verified is hasattr-guarded (init_command.py:1759, :1760, :1768) and _ctx_warning is still initialized in __init__, so removing the __init__ seed introduces no AttributeError risk on any other code path.

Issues Found

🟢 Minor — tests mock _test_model_inference rather than exercising the real SD path (test_init_verify_ctx_display.py:94)

The tests drive the display logic by stubbing _test_model_inference and setting cmd._ctx_verified directly, which is the right way to test the display bug. But it means the suite won't catch a future regression where the real SD/embedding branch (init_command.py:1585) accidentally sets _ctx_verified. The test comments honestly call out this assumption ("SD path in _test_model_inference never touches _ctx_verified"), so this is a note, not a blocker — consider one integration-style test that lets the real _test_model_inference run against a mocked client for an SD model, if you want to lock that invariant down too.

Strengths

• Root cause, not symptom. The fix removes the stale-state source (__init__ seed) and centralizes the per-model reset at the loop top, rather than patching the display site. The now-redundant post-display delattr is correctly removed.
• Comments follow the house style — they explain the why (the stale-flag inheritance invariant) in one place each, exactly where a future reader needs it, without narrating the mechanics.
• Good test coverage of the actual scenarios: SD-first stays clean, LLM-without-reported-ctx still warns, and ctx state doesn't leak across models (with forced ordering). The third test asserting on the specific SDXL-Turbo line rather than global text is a nice touch.

Verdict

Approve. No blocking issues; the lone minor note is optional follow-up. Bug fix is correct, scope-clean, and tested — verified passing locally. Docs need no update (console-output fix, no CLI/API surface change).