feat: gpt caching

[schiwekM]

Prompt caching is now also applied to GPT models.

Have you...

• Added relevant entry to the change log?

[Akatuoro]

AI review:

lib/models/aicore.js
  ● 174 [correctness] Tool `cache_control` injection for Claude is removed with no replacement, silently disabling tool-definition caching.
  ● 126 [correctness] `mergeOrchestrationConfig` accesses `config.promptTemplating.model` without null-checking `config.promptTemplating`, crashing if super returns a config without that field.
lib/models/aicore-caching.js
  ●  36 [correctness] `withPromptCachingOptions` replaces multi-breakpoint per-message injection with a single top-level `cache_control`, reducing Claude caching coverage to one message instead of up to four.
  ● 100 [correctness] `_cacheKeyPart` truncates to 32 chars, causing `prompt_cache_key` collisions between Claude 3.5 Sonnet versions deployed simultaneously in AI Core.
  ●  59 [correctness] `buildPromptCacheKey` uses `user || thread`, so a `null` or empty-string `userId` silently falls through to `thread_id`, producing the wrong affinity bucket.
  ●  88 [efficiency]  `isGptExtendedRetentionModel` and `isGpt56OrLater` are both called without short-circuit in `isPromptCachingModel`, computing `_gptVersion` twice per call.
lib/telemetry/chat-tracing.js
  ● 169 [correctness] `_detectCacheRequest` checks `modelOptions.params.prompt_cache_key` but the cache key is never stored in model params at construction time, making that branch permanently dead.

The selection logic seems quite complicated, as evidenced in the bullet points above... can we boil it down to easier criteria?
Also, aicore-caching.js should not be located in lib/models, that is reserved for the impls.

[Akatuoro]

Works for anthropic provider. For aicore provider, it works with gpt-5.5 and anthropic models. For other gpt models (gpt-5.2, gpt-5.6-sol, ...), I get Error while iterating over SSE stream.

Also for aicore, I see these in the log:

[agents] - [mlflow] GET /api/2.0/mlflow/registered-models/get?name=srv%2Ftravel-agent%2FAGENTS.md → 400: {"error_code": "INVALID_PARAMETER_VALUE", "message": "Invalid model name 'srv/travel-agent/AGENTS.md'. Names cannot contain '/' or ':'.", "sqlstate": "KAM00", "error_class": "INVALID_PARAMETER_VALUE"}
[agents] - [mlflow] POST /api/2.0/mlflow/registered-models/create → 400: {"error_code": "INVALID_PARAMETER_VALUE", "message": "Invalid model name 'srv/travel-agent/AGENTS.md'. Names cannot contain '/' or ':'.", "sqlstate": "KAM00", "error_class": "INVALID_PARAMETER_VALUE"}
[agents] - [mlflow] POST /api/2.0/mlflow/model-versions/create → 400: {"error_code": "INVALID_PARAMETER_VALUE", "message": "Invalid model name 'srv/travel-agent/AGENTS.md'. Names cannot contain '/' or ':'.", "sqlstate": "KAM00", "error_class": "INVALID_PARAMETER_VALUE"}

[schiwekM]

The mlflow errors are fixed in the other PR. gpt 5.2 is not available with version latest in AI core causing the error and got-5.6-sol did not accept temperature = 0.
For the first scenario I am adjusting the AI Core class that the errors are caught and correctly thrown avoiding the generic error message.
For 5.6-sol I am adding a condition that models we know don't accept a temperature of zero do not get it.

[Akatuoro]

Setting temperature to 1 ourselves is a bad idea. We should ideally not have to provide it at all

[schiwekM]

Fair point. Deleting it should work as well and I switched to that now.