My guess is Roux it depends. (Yes, even if we restrict to face turn metric. I don't care much for STM either. Everything else in this post will use FTM. I'll explain why I think Roux is generally better later.)
"CFOP" and "Roux" are not methods so rigorously defined that you could ask two different people to code implementations of these methods, without further instructions, and end up with solvers that always produce the same solutions. I'm much more familiar with CFOP than with Roux, so I'll use that as an example first.
CFOP is just cross, F2L, OLL, then PLL, right? Right? Wrong.
Very often, you'll have multiple choices of cross colour that take the same number of moves. Which do you choose? Do you pick one arbitrarily, or do you pick one that has the best continuation?
That's a trick question; you don't pick cross colours for their continuations, you pick cross solutions. Do you generate all cross solutions, then pick the one that has the best first pair? What if e.g. white cross has a 5-move solution that leads into 3-move first pair, while yellow cross has a 4-move solution but 7-move first pair?
Moving on to F2L, do you always pick the best pair to solve? What about multislotting? Are algs that exploit free slots allowed? Are adjacent pairs prioritised over diagonal pairs, and if so, to what extent?
OLL and PLL are the mechanical parts involving no degree of freedom, and those aren't really up for dispute.
---
You know where this is going already, but just for the sake of formality…
Roux is just FB, SB, CMLL, then LSE, right? Right? Wrong.
FB runs into the same CN-related issue as above.
SB and CMLL have no degree of freedom (assuming we're not using alt solutions for better continuations).
LSE as done by beginners is usually broken into three steps: EO ("4a"), UL UR ("4b"), L4E ("4c"). As done by experts, EO and UL UR are (partially) combined into a single step, EOLR. (EOLR usually involves setting up EO to the arrow case first, but there's also full EOLR which solves 4a and 4b directly.)
If we're to compare methods of similar alg counts, it would only be fair to include EOLR as part of Roux.
(edit As GodCubing pointed out, I misunderstood what EOLR is. This does not materially affect the numbers presented here; EOLR proper uses more algs, but also has slightly lower move count than the variant I described above.)
Also, since we're using FTM as the metric of choice, of course we'll also be using solutions for the individual steps that are optimal in FTM, not STM.
---
Let's say we solve the steps individually, without regard for continuations.
CFOP is, in reality, a seven-step method:
1. Cross
2. First pair
3. Second pair
4. Third pair
5. Last pair
6. OLL
7. PLL (and AUF)
Roux is only a six-step method:
1. First block
2. Second block
3. CMLL
4. EOLR setup (either set up to arrow or solve 4a)
5. EOLR
6. 4c
Assuming we allow using free slots, the worst case for each F2L pair is at least 8 moves (corner twisted in place and edge flipped in place), except for the very last pair having a worst case of 9 moves (corner solved, edge flipped in place). That makes F2L take 8+8+8+8+9 = 41 moves. OLL's worst case is 12 moves (OLL 20), while PLL's worst case is 14 moves (V perm at any AUF). This gives a total of 41+12+14 = 67 moves.
I don't know the exact numbers for Roux, but my guesstimates are:
FB: 9
SB: 12
CMLL: 11 (known to be the worst case)
EOLR setup: 8
EOLR: 13
L4E: 8
Total: 61
There's enough slack that I could be slightly wrong in a few of these and Roux's worst case would still be better than CFOP's worst case.
I imagine that if you restrict to purely MU algs for LSE, or if LSE has to be done as 4a-4b-4c, Roux would lose by a significant margin, but for reasons I already stated above, these would be silly restrictions to impose.
