fix(ai): assume newer models keep family features

[nexxeln]

Follow-up to #51339 | #48513

Why the change

Features like mid-conversation effort switching were turned on by exact model id, so every new launch (gpt-6.1-astra, claude-sonnet-5-1, …) lost them until we shipped a code change; now newer versions keep what their family already supports, with a config override for the exceptions.

Special things to note

• A wrong guess now fails louder. Sending configuration_update or the Anthropic effort beta to a model that doesn't support it returns a 400 instead of just missing the prompt cache. Newly on but unverified: gpt-6-*-fast, dated GPT-6 snapshots, Sonnet/Haiku 5.1+, and new Claude families. The fix for any of them is compatibility.supportsEffortUpdates: false.
• GPT-6+ models from OpenAI.configure().responses() now get the same default reasoning: { effort: "medium", summary: "auto" } as GPT-5. Before this they got no reasoning defaults at all.
• models.dev can't set the new overrides yet because its schema has no field for them; only user config and plugins can. plugin/verbosity.ts still skips gpt-7 on purpose, so it's unchanged.

Change outline

The rules, before and after:

 OpenAI effort updates                       ai/src/protocols/openai-responses.ts
-  id is exactly gpt-6-astra | gpt-6-sol | gpt-6-luna
+  gpt-N with N >= 6, except -pro slugs
   still off: reasoning.mode "pro", context_management, Bedrock Mantle `openai.` ids

 OpenAI facade reasoning defaults            ai/src/providers/openai-options.ts
-  id includes "gpt-5"
+  gpt-N with N >= 5
   still off: gpt-5-chat, gpt-5-pro

 Anthropic effort updates                    ai/src/protocols/anthropic-messages.ts
-  Opus >= 5, Fable/Mythos >= 5.1, never Sonnet, Haiku, or other families
+  any Claude family >= 5.1, plus Opus 5

 Anthropic mid-conversation system messages  ai/src/protocols/anthropic-messages.ts
-  opus | sonnet | haiku | fable | mythos >= 5, plus Opus 4.8
+  any Claude family >= 5, plus Opus 4.8

 Copilot Responses reasoning detection       core/src/github-copilot/responses/openai-responses-language-model.ts
-  id starts with "gpt-5"
+  gpt-N with N >= 5

The Copilot one was the same AI SDK problem: on Copilot, gpt-6-* silently dropped reasoning.effort, kept temperature, and sent the system prompt as system instead of developer.

The override already existed in @opencode/ai, but config silently stripped these keys when decoding. They're now part of the shared contract:

 // packages/schema/src/model.ts
 Model.Compatibility {
   ...
   supportsPromptCacheKey?: boolean
+  supportsEffortUpdates?: boolean
+  supportsThinkingBlockBinding?: boolean
 }

How a config override reaches the request:

provider.<id>.models.<id>.compatibility   (user config or plugin transform)
  → Model.Info.compatibility
  → ModelResolver.fromCatalogModel        merges it onto LanguageModel.compatibility
  → supportsEffortUpdates(request)        the override wins over the id guess

"provider": { "openai": { "models": { "gpt-6-sol-fast": { "compatibility": { "supportsEffortUpdates": false } } } } }