fix: title generation fails with OpenAI reasoning models

[trungutt]

• Title generation with OpenAI reasoning models (gpt-5, o-series) fails with empty title output from model because these models always reason internally, even when no reasoning configuration is provided. The hidden reasoning tokens consume the entire output token budget, leaving nothing for the visible title text.

• When the code intended to "skip reasoning", it simply omitted the reasoning parameter from the API request. This does not disable reasoning - the model defaults to its standard reasoning behavior, which can easily exhaust a small token budget.

• The output token budget (20 tokens) was sized for non-reasoning models and far too small once reasoning overhead is factored in.

Fix

• Explicitly request the lowest reasoning effort supported across all OpenAI reasoning models so the model minimizes internal reasoning token usage.
• Increase the output token budget to accommodate the reasoning overhead that reasoning models always incur.

How to test

mise build

# The original failing command from the issue
./bin/docker-agent debug title ./examples/rag.yaml "What can you do?"

# Explicitly test reasoning models
./bin/docker-agent debug title ./examples/rag.yaml --model=openai/gpt-5 "What can you do?"
./bin/docker-agent debug title ./examples/rag.yaml --model=openai/o3-mini "What can you do?"
./bin/docker-agent debug title ./examples/rag.yaml --model=openai/o4-mini "What can you do?"

# Non-reasoning models should still work as before
./bin/docker-agent debug title ./examples/rag.yaml --model=openai/gpt-4o "What can you do?"
./bin/docker-agent debug title ./examples/rag.yaml --model=anthropic/claude-haiku-4-5 "What can you do?"

Each command should print a short title to stdout without errors.

Fixes #2318