Facts and figures

Optimising variables for wall-hugging
Server framerate	Best angle with `cl_fs` = 200^a	Best angle with `cl_fs` = 400^b	Best `cl_fs` when parallel^c	Maximum possible speed
(fps)	(degrees)^d		(pixels/s)
10^e	24.9	7.0	500	326.6
11	19.7	1.4	390	332.7
12	15.5	-2.8	340	339.4
13	12.7	-7.0	310	346.7
14	9.8	-9.8	280	354.2
15	7.0	-11.2	260	361.9
16	5.6	-14.1	250	369.5
17	4.2	-16.9	230	377.2
18	2.8	-16.9	220	384.9
19	1.4	-18.3	210	392.5
20^f	1.4	-18.3	203	400.0
21	1.4	-18.3	207	407.0
22	2.8	-16.9	211	413.1
23	2.8	-16.9	215	418.7
24	2.8	-16.9	219	423.5
25	2.8	-16.9	222	427.6
26	2.8	-15.5	225	431.6
27	4.2	-15.5	227	435.0
28	4.2	-15.5	229	438.2
29	4.2	-15.5	232	441.2
30^g	4.2	-15.5	234	443.8
31	4.2	-14.1	235	446.2
32	4.2	-14.1	237	448.4
33	4.2	-14.1	239	450.6
34	5.6	-14.1	240	452.6
35	5.6	-14.1	242	454.5
36	5.6	-14.1	243	456.2
37	5.6	-14.1	244	457.7
38	5.6	-14.1	246	459.2
39	5.6	-14.1	247	460.4
40	5.6	-14.1	248	461.8
41	5.6	-12.7	249	463.2
42	5.6	-12.7	250	464.3
43	5.6	-12.7	251	465.4
44	5.6	-12.7	251	466.5
45	5.6	-12.7	252	467.5
46	5.6	-12.7	253	468.5
47	7.0	-12.7	254	469.5
48	7.0	-12.7	254	470.2
49	7.0	-12.7	255	471.0
50	7.0	-12.7	256	471.8

Notes

a: This is the angle to the wall at which you should run if you play with "Always Run" off and do not alter cl_forwardspeed.
b: This is the angle to the wall at which you should run if you play with "Always Run" on and do not alter cl_forwardspeed.
c: This is the value you should set for cl_forwardspeed if you run parallel to the wall
d: A positive angle is out from the wall, a negative angle is in to the wall
e: NQ servers with sys_ticrate 0.1 have this fps. As you can see, wall-hugging isn't worth bothering with on these.
f: NQ servers with sys_ticrate 0.05 (the default) have this fps. Here you get useful results.
g: Most QW servers run at least as fast as this, so wall-hugging is certainly very effective on these servers.

Getting these results involved all kinds of messing around including building some bot-like players that would run the experiments and using them to collect thousands of sets of data. I found it fun, but I think there is probably something seriously wrong with me. People who helped with this "cyberscience" by running tests or writing utilities include Stefan Schwoon, Nolan Pflug, Yonatan Donner, Patrick Cupka, Drastic_Man, and Chris Longden - my thanks to them all.