A prototype proves the circuit. It does not prove that the board can be assembled at volume, panelised without damage, or tested in under a minute. Those are separate problems, and they are decided during layout, by which point fixing them is a respin.
Ask the assembler before you route
Every assembler has a process window: minimum component spacing, preferred paste thickness, which packages they can place reliably, what they want for fiducials, and how they like boards panelised. Ask for their DFM guidelines and design to them. It costs nothing at the start and saves an argument later.
The recurring problems are mundane. Components too close to the board edge for the conveyor rails. No fiducials, or fiducials too close to the edge. A fine-pitch part next to a large thermal mass, so one or the other will reflow badly. Tall components adjacent to a tab-routed breakout. Polarity markings hidden under the part they mark, which turns into an inspection problem on every unit built.
Design the test fixture while you design the board
This is the part that gets left out, and it is where the cost sits at volume. If you want to test in-circuit, you need test points: proper pads, large enough for a pogo pin, on one side of the board, with keep-outs around them, and a couple of tooling holes so the fixture locates repeatably. Retrofitting that is another respin.
Think about what you actually want the fixture to prove: power rails within tolerance, the MCU programmable, the radio transmitting at expected power, sensors responding, and a unique identity or key programmed and recorded. That list determines the test points, the connector, and how long each unit spends in the fixture. A minute per unit on a ten-thousand-unit run is a hundred and sixty hours of someone's time.
The specific checks worth running
Before release to fabrication: run the DRC against the fabricator's real capability rather than the default; check the drill-to-copper and annular ring against their class; confirm the stack-up is one they actually build rather than an idealised one; verify impedance-controlled traces are calculated for that stack-up; check paste layer apertures on QFN and BGA thermal pads; confirm the panelisation, with tabs or v-score placed where they will not stress components; and check that every part in the BOM has a real, orderable part number with stock, not a generic.
On that last point: run the BOM for lifecycle status and lead time before you commit the layout. A part that goes end-of-life between prototype and production means a respin at the worst possible moment, and it is entirely foreseeable.
Where this fits
None of this is glamorous and none of it is hard. It is a checklist run by someone who has had the phone call from an assembler at 7am, and it turns a board that works into a board that can be built. If you are about to release files and nobody has run that pass, it is worth a few hours before you spend the fabrication money.