TL;DR:
I wonder if that observation is heavily biased by looking at today's all-influencing x86 architecture and its direct predecessors (x80). From my memory the described behaviour is only shown by a small minority of CPU designs. That is unless 'most' is measured by sold units (*1).
Details
Which CPU was the first to clear the carry and overflow bits after performing logical operations?
Asking for a first only makes sense if there is a linage (see below) but not in general.
notice over and over how consistently most 1980's CPUs echo a pattern.
Well, no designer lives on an island.
Since an overflow or carry cannot result from a logical operation like AND or OR, it seems like the designers could have chosen to design the CPU to not change the "carry" and "overflow" bits.
Which is what many did.
However what I'm seeing almost universally is that whenever a logical operation like AND or OR is performed, the "carry" flag and the "overflow" status flags are cleared, while the "sign" or "negative" and "zero" flags are set according to the result of the operation.
It may be very helpful if you could add some information on which you base this claim of being 'almost universal' as I can come up with many CPUs not clearing one or both of them.
If possible with a separation between original designs and designs that inherited that behaviour due to the need for compatibility.
I remember, from my old Z80 programming days, using this side effect to clear the carry flag with an and a instruction; this was already common practice back in the 1980's.
Not sure why this should be called 'common practice' as it's the explicit behaviour of the Z80, which inherited it, due to the requirement of full compatibility of carry handling from the 8080, and it in turn from the 8008. Thus the Z80 couldn't introduce an OV of its own but had to repurpose the Parity flag. And in turn the 8086 and all its offspring also had to work that way. No freedom of choice here.
When was this convention of clearing the carry and overflow flags after a logical operation introduced,
That's the first time I'm hearing of that being a convention.
how deliberate was it,
While there are many independent designs, not all had the freedom to decide - especially not the ones that were all about preserving compatibility, foremost visible here the 8008/8080/Z80/8086 lineage.
and which CPU architecture implemented it?
Since you already state that a majority according to your own research does it, it might be more useful to ask which didn't.
The World According to Flags
I can right from the top of my mind come up with a whole lot of CPUs with different behaviour:
No change at all, except arithmetic (i.e. non logical):
Partial change outside of arithmetic (in most cases clearing OV leaving C unchanged)
- Zilog Z8 - All ALU operations, i.e including logical (*4)
- Motorola 68xx - All instructions loading a register (that is including logical)
- PDP-11 - Most instructions clear OV
- PDP-11 - All shift set OV to Carry XOR Negative (*3)
And then there are those that update all flags with all ALU operations with the result of that operation - or a default value where that flag has no meaningful value. Usually 0.
- Motorola 68000 - Clear both
- Z80/8086/etc. - Clears both (*4)
Interesting that the two most dominant microprocessors of the 1980s fall into that group... that wouldn't by accident create a bias, or would it?
Then there are similar ones but without an Overflow Flag, so only Carry gets modified with each ALU operation
And there you also got one lineage coming from a first, Datapoint 2200 clearing Carry with logic instructions and going all the way to today's (somewhat) most used designs.
And then there's the odd bunch:
- Transputer - A single Error Flag does everything :))
- NS 16/32xxx - Logical operations change no flags at all
Looking below
Especially the (in)famous PDP-11 gives insight about a basic difference between
- machines that produce a full or partial condition code and
- machines that handle and update status flags.
The first is a class where each and every (non-special) instruction results in a new condition code, which may or may not include previous information. The IBM /360 might be the most clear and well known. Here instructions do not set flags but result in an abstract condition code in form of a single value (0..3). The PDP-11 is also a condition-code machine, but here it's represented as 4 bits which may be handled separately depending on instruction.
The second class handles flags as single bits and updates them independently according to instruction and only when either bit is meaningful to that instruction - except they are so simple that they update it always after an ALU operation.
*1 - Even then, the answer may still be some other family hidden in anything from toothbrushes all the way to forklifts :))
*2 - Which funnily enough only got Carry and Overflow as fixed ALU flags :))
*3 - It might be established that I'm not exactly a fan of the PDP-11, but that little piece always intrigued me. Noting a sign change can be quite handy after some operations.
*4 - The Z8 behaving differently from the 8080/Z80 is a nice hint that at least one of the 8080 creators wasn't all that happy with setting both by all ALU operations.