Added a per-agent consult level in core (0 to 3, registered at the composition root) so each companion can be told how much to think versus just act. The wedding wires the budgeter to deep, advisor/designer/scheduler to standard, and supplies the actual consultative prose in the vertical since what "expert" sounds like differs between software and a wedding. Chat approval messages now render as task cards with real Yes / Not-now buttons hitting a grant-gated work-item endpoint, with the buttons suppressed once the item has settled so a page reload does not re-show live controls on an already-approved card. Also ripped the agency-specific wording ("client of their software team", "ship it") out of the shared composer, taught the budgeter to actually persist the stated total before allocating (the from-scratch $0 bug), added human describe() for facts so the approval card stops showing raw budget.drift JSON, and threaded a seam trace through the LLM loops for some visibility into what the agents are doing.

5 commits across 1 repo (belo-works: 5). 27 files changed. Diff was truncated for summarization.