Regarding the problem with shaping vlan-tagged traffic on linux bridge — it was my fault :-)
I asked firstname.lastname@example.org, and Sergey Vlasov answered me with some info and hint — thanks a lot!-)
As Sergey told, «a packet is being passed to bridge module before 8021q can process it». That’s why the classifier decides that a packet does not match a filter rule — because filter contains «protocol ip» condition, however a packet contains «protocol 802.1q» actually. It look like it’s enough to use «protocol 802.1q» in filter condition — after that ip addresses (offsets from IP packet’s start) will be counted correctly.
This offers a possibility to build «aggregating shapers» on linux bridges, which will be able to shaper clients” traffic regardless of «traffic direction» (regardless of a particular vlan membership). And it will be very easy to migrate from an «usual» shaper on linux bridge — we will have to replace «protocol ip» with «protocol 802.1q» in a filter template (or we can add one more filter — if it is really necessary).