Decoupling caps placed right next to the IC but still seeing power rail noise

If individual cap placement is genuinely correct, the next thing I'd check is the ground via itself, not the capacitor. A decoupling cap with a perfect loop on paper but a single ground via shared with several other components, or a via that lands on a ground plane with a gap or split nearby, still has more loop inductance than it looks like from the placement alone.

A few specific things to check:

• Does each decoupling cap have its own dedicated via to ground, or are several caps sharing one via through a short trace? Sharing reintroduces inductance you thought you'd eliminated by placing the cap close in the first place.
• Is the ground plane actually solid and continuous directly under that section, or is there a gap, slot, or another via field interrupting it nearby? (Worth checking after any later placement change — see the 4-layer stack-up thread for how easy this is to introduce without noticing.)
• Are you decoupling with only the one value the datasheet specifies, or do the active peripherals draw current at a frequency range the single cap value doesn't cover well? A second, smaller-value cap in parallel sometimes recovers margin a single value leaves on the table.

Datasheet-correct value and "looks close on the layout" get you most of the way, but the via and plane continuity is where the remaining noise margin usually hides.