Live assistant pass: verify draft-a-block end to end (now adopts module suggestions)

[ApocDev]

The assistant's draft-a-block flow has been through two refactors of its module handling without a live run: it first harvested ambient auto-fill results, and now solves once, adopts the per-row suggested fills as the draft's explicit module picks (pinning the machine they were sized for), and re-solves so counts/power/imports reflect them (buildBlockDraft in app/src/server/agent-tools.server.ts).

Types and unit tests pass but nothing exercises the path against a real model. Needs a live OpenRouter session: draft a block, confirm the draft arrives with sensible modules (prod where allowed, speed/efficiency split elsewhere), machines pinned to match, and rates consistent with the filled solve. General live pass on the other assistant flows (plan drafting, revise-rate) is also still outstanding since the solver v2 and suggestion changes landed underneath them.

[ApocDev]

Verified the deterministic layer against real project data (block 27's end-game iron chain, 23 recipes at 400/s iron plate), by calling the submitBlock tool's execute directly — the same path the model invokes, minus the LLM:

• 21 of 23 rows arrive module-filled (the other two run in slot-less machines), each with its machine pinned.
• The draft's power figure exactly matches an independent re-solve using the draft's own module/machine picks, so the reported rates/imports/byproducts come from the filled solve, not the provisional one.
• Re-solving with the adopted picks leaves zero rows still differing from their suggestion — adoption is a fixpoint, including the burner rows.
• Imports and byproducts look right for the chain (borax/coal/crude-oil/… in, ash/gravel/stone/tailings out).

What this does not cover: the model actually choosing recipes and calling the tool sensibly, plan drafting, and revise-rate — those still need the live OpenRouter session, so leaving this open for that.