fix(responses): merge leading system messages in user_note_safe fallback

[imi4u36d]

Problem

When Codex App Desktop sends /v1/responses requests with both instructions and a system/developer message in input, two leading system messages are produced. If the conversation also contains a mid-system message (e.g. via previous_response_id chain or multi-turn input), the user_note_safe downgrade path in prepare_system_messages_for_template preserves the leading system block as-is without merging.

Since _merge_consecutive_roles only merges user/assistant roles, the two consecutive system messages survive into the chat template. Strict templates like Qwen3.6 validate that system messages can only appear at the first position, causing:

POST /v1/responses -> 400: Chat template error: System message must be at the beginning.

Root Cause

In omlx/api/utils.py, _downgrade_mid_system_to_user_notes preserves leading system blocks via rewritten.extend(messages[start:i]) at line 601-602. The caller unsupported_fallback then only applies _merge_consecutive_roles, which doesn't touch system messages.

Fix

Call _merge_consecutive_system_messages in unsupported_fallback before _merge_consecutive_roles when the user_note_safe path succeeds. This merges consecutive leading system messages into one, satisfying strict template validation.

Reproduction

from omlx.api.utils import prepare_system_messages_for_template

class Qwen36Tokenizer:
    def apply_chat_template(self, messages, **kw):
        for i, msg in enumerate(messages):
            if msg.get('role') == 'system' and i > 0:
                raise Exception('System message must be at the beginning.')
        return 'ok'

messages = [
    {"role": "system", "content": "You are a helpful assistant"},  # from instructions
    {"role": "system", "content": "Be concise"},                   # from input
    {"role": "user", "content": "Hello"},
    {"role": "system", "content": "Remember this"},                # mid-system
    {"role": "user", "content": "Continue"},
]

result = prepare_system_messages_for_template(
    messages, Qwen36Tokenizer(), unsupported_mid_system_policy="user_note_safe")
# Before fix: ['system', 'system', 'user'] → template error
# After fix:  ['system', 'user']           → OK

Closes #1908

[jundot]

Thanks for the fix. The change is scoped to the user_note_safe fallback and uses the existing system-message merge path after mid-system notes are downgraded, so it addresses #1908 without changing the native-preserve path. I verified the API utility tests locally and CI is green. This looks good to me, and I'm going to merge it.