Feedback deploy rides along dont fragment

When a feature-build agent (Opus/Sonnet) finishes reasoning work, merges to main, and needs to deploy, it should deploy right there in the same session rather than spinning up a new Haiku session just for the deploy command.

Why: Justin pushed back on the idea of splitting every mechanical step to a cheap model/session. A trivial command like deploy.sh produces almost no output, so it costs nearly nothing regardless of model, the expense of a frontier model is in thinking and writing a lot, not firing one command. Starting a fresh session to run one command costs more (session boot + re-explaining context) than letting the agent that did the work finish the job. Confirmed by Justin's own reasoning in-session ('I guess when it gets merged to main you still need to deploy the site... is it really worth starting a new session just to do that') and validated as correct.

How to apply: Only spin a fresh cheap-model session for tasks that are standalone and mechanical from the start with no preceding reasoning (e.g. 'deploy current main', 'run the backup', 'push these staged files' as the entire point of the session). Do not fragment a deploy/merge step off the end of a feature-build agent's own work.

[auto-memory session 18425db8-578a-4927-8d0a-0a385a164ba3, confidence 0.85, mode direct]