Discovery agent improve

[dimavrem22]

• window properties are used for routine creation
• improved routine integrity (automatically apply escaped quotes when necessary, fixed validation bugs) [fully tested in unit tests]
• openai responses api parsing
• better prompts
• slight data model updates to lm response classes
• tiny updates to network monitoring (more supported file types + .txt is default response body file type)
• should be using GPT 5.1 (much faster and feels a bit smarter)
• responses api appears to have sufficient context to run for extended period of time

Issues addressed:

• Use OpenAI responses structured output #51
• Routine Discovery shouldn't produce unused parameters #46

Future PRs will address

• run the generated routine and evaluate success before finishing #57
• manually interpolate all placeholders to log errors in logic #58