fix: keep developer messages in DeepSeek-V4 thinking drop

[longzhenren]

Root cause

_drop_thinking_messages in deepseek_v4_encoding.py documents that "Developer messages before the last user are dropped entirely", and its keep_roles set indeed lacks developer. But the rest of the module supports the role: the renderer has a developer branch (rendered like a user turn) and find_last_user_index counts it.

So in thinking mode with drop_thinking=True (the default), a developer message before the last user turn is silently removed from the prompt: the instructions never reach the model, and encoding a conversation with or without the developer message produces the same output.

Reproduction:

from lmdeploy.deepseek_v4_encoding import encode_messages

base = [
    {'role': 'system', 'content': 'You are helpful.'},
    {'role': 'user', 'content': 'Real question'},
]
with_dev = [base[0], {'role': 'developer', 'content': 'Always answer in French.'}, base[1]]

a = encode_messages(base, thinking_mode='thinking')
b = encode_messages(with_dev, thinking_mode='thinking')
assert 'Always answer in French.' not in b  # developer instructions silently lost

Fix

Add developer to keep_roles and update the docstring, consistent with how the renderer and find_last_user_index already treat the role.

Test

• test_deepseek_v4_thinking_drop_keeps_developer_message: [system, developer, user] in thinking mode renders the developer content.
• test_deepseek_v4_thinking_drop_keeps_developer_in_multiturn: a developer message before the last user turn survives the drop while assistant reasoning_content is still stripped.
• Both tests fail on the current code and pass with the fix; the existing test file still passes (14 passed).

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.