The quarter-wave rule is the right frame, but I would add two measurement-side points and one process-side point, because the rule alone tends to produce a number that nobody can verify on the finished board.
First, the notch frequency is the checkable artifact, not the stub length. For a stub of length L with in-plane velocity v, the first null sits at f = v / (4L). Working backwards from a measured S21 null is far more reliable than trying to section a 0.3 mm residual: a 0.5 mm stub lands near 72 GHz, 0.65 mm near 56 GHz, and 1.3 mm near 28 GHz. That means for 112G you can accept a link if the S21 notch stays comfortably above the band you actually use, and you do not need a microsection to say so. Conversely, a coupon that looks clean in section but shows a shallow null just above Nyquist is telling you the residual is marginal.
Second, stub length is not the only term. The via barrel's excess capacitance and the antipad diameter both shape how deep that notch is. A long stub on a via with a large antipad couples less energy into the stub and shows a shallower null than a short stub on a tightly coupled via. So "backdrill every via longer than 0.5 mm" can be simultaneously over-tight in one place and under-tight in another. The useful spec is a return-loss or insertion-loss deviation budget on the critical pairs, with the stub length as the manufacturing proxy - not the other way round.
On measurement limits: TDR resolution is set by the edge, and a 10-15 ps edge is roughly millimetre-scale, so TDR will not resolve a 0.3 mm target on a real board. Use TDR for gross defects (nicked target layer, double-drill) and use coupon S21 plus a per-panel microsection for the depth number. Asking a fab to prove 0.3 mm with TDR invites a number that means nothing.
Process questions worth putting in the PO, since depth tolerance is where the money and the opens are:
- Target depth and tolerance as separate line items (+/-0.05 vs +/-0.1 mm on a 0.3 mm target is a 30% difference in residual).
- Whether backdrill happens before or after final plating, and how the barrel is re-planarised - an unplanarised stub end still resonates, and residue in the hole affects the next lamination.
- Registration and drill wander statistics per lot, not just capability on the best panel.
- Coupon plan: one impedance coupon per panel with a microsection of the deepest residual via, plus S21 on the same coupon.
- Acceptance on the notch: define the highest frequency you will use and require the null to sit above it with margin, rather than requiring a stub number you cannot measure.
One more practical note on scope: only the escapes on the pairs that actually run at the top rate need this. On a mixed board, that is usually a small fraction of the via count, and putting backdrill on everything is how the cost objection wins the argument and the critical pairs end up under-specified. Split the stackup drawing into "critical" and "general" via classes and price them separately.