Drop gpt-4.1-nano and o4-mini ahead of their Oct 23 shutdown; rebuild README model list

[adambalogh]

Follow-up to #309, which merged before this last commit landed on the branch. Rebased onto current main.

Removed: GPT_4_1_NANO and O4_MINI

OpenAI shuts both down on 2026-10-23, ~1 month out (deprecations).

These are unlike the gpt-5 / o3 entries. Those are undated aliases whose dated snapshot retires; here OpenAI's shutdown list names the exact bare slugs gpt-4.1-nano and o4-mini — which is precisely what the gateway sends — so they hard-fail on the date. An SDK release plus user upgrade lag outruns 32 days, so shipping them ships a known break. Replacements are OpenAI's own: GPT_5_6_LUNA and GPT_5_6_TERRA, both already in the enum.

Documented but kept: O3, GPT_5, GPT_5_MINI

These are undated aliases of snapshots retiring 2026-12-11 (o3-2025-04-16, gpt-5-2025-08-07, gpt-5-mini-2025-08-07). OpenAI's deprecations page does not state whether such an alias is repointed to a successor or retired alongside its snapshot.

If it's repointed, that's the same silent-redirect trap the retired Grok slugs had — requests keep succeeding while quietly running a different model at a different cost, which is exactly what went unnoticed for four months on the xAI side. Noted in the enum so it gets revisited before December.

Rebuilt README's "Available Models" section

This list used prose names (- GPT-5 Mini) rather than model ids, so the greps in #309 didn't catch it. It had drifted badly:

• listed only six OpenAI models, stopping at GPT-5.2
• stopped at Claude Opus 4.8 and Gemini 3.5 Flash
• named the retired Grok 4, Grok 4 Fast, Grok 4.1 Fast
• omitted ByteDance, Tencent and Z.ai entirely

It's now generated from the TEE_LLM enum and lists all 63 models with their gateway ids and enum members, grouped by provider, with image-generation models marked. Prose and code can no longer drift apart.

Verification

• 63 enum members, every one resolves against the gateway's _MODEL_LOOKUP (checked by importing the registry)
• make test — 179 passed
• make check — ruff clean, mypy clean on src (30 files) and examples (16 files)
• Docs regenerated with make docs

Still open (not in this PR)

Unchanged from #309 — the SDK is behind the gateway on more than the model list:

• llm.chat() has no aspect_ratio or resolution parameter, so Seedream 5.0's cheaper 1.5K tier and image shape selection are unreachable from the SDK.
• No client for the /v1/web_search endpoint.
• Nothing surfaces the gateway's moderation response block.
• The web_search=True docstring still describes it as "the provider's native web search"; the gateway treats that flag as a deprecated no-op.

The gateway-side mispricing flagged in #309 is now fixed in OpenGradient/tee-gateway#166.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QmnaiLePKAQEkNfR8VPfja

---

Generated by Claude Code