fix(llms): extend _is_reasoning_model to cover o4 family and dated Azure subversions

[charlesaurav13]

Linked Issue

Closes #5296

Description

_is_reasoning_model matched exact names (o1, o3, gpt-5, …) and a short prefix list (o1-, o3-). Azure OpenAI surfaces these models under dated deployment names such as:

Deployment name	Expected result
gpt-5-2025-06-15	reasoning ✗ (missed)
gpt-5.4-2025-07-15	reasoning ✗ (missed)
gpt-5.4-mini-2025-07-15	reasoning ✗ (missed)
o4-mini-2025-04-16	reasoning ✗ (missed — o4 family absent entirely)

Any missed model passes max_tokens to the API and receives a 400 rejection.

Changes

• Added o4 and o4-mini to the exact-match set (new reasoning model family)
• Extended the prefix tuple with o4-, o4., gpt-5-, gpt-5. to catch all dated and point-release subversions
• Removed the misleading inline comment claiming gpt-5.x supports temperature

# Before
reasoning_prefixes = ["o1-", "o1.", "o3-", "o3."]

# After
reasoning_prefixes = ("o1-", "o1.", "o3-", "o3.", "o4-", "o4.", "gpt-5-", "gpt-5.")

Type of Change

• Bug fix (non-breaking change that fixes an issue)

Breaking Changes

N/A — only broadens the set of models treated as reasoning models.

Test Coverage

• No tests needed — pure name-matching logic, deterministic and environment-independent

[kartik-mem0]

Hi @charlesaurav13 — thanks for working on this! We've just merged #5327 which addresses issue #5296 by adding an explicit is_reasoning_model config override rather than extending the name-based heuristic. Since a broader prefix match would break the OpenAI gpt-5.x non-reasoning contract (issue #4738), the opt-in flag approach was the safer path forward.

Closing this PR as the issue is now resolved, but we appreciate the effort — feel free to pick up another issue!