Note/preamble: I'm not sure why I got more DVs for this answer, but do note that I tried to answer the Q as asked namely
When worn by members of the public who are infected, are surgical masks effective at slowing the spread of COVID-19 ("source control")?
Answering with population studies on mask mandates in general, would have been really easy and given a nice what-you-probably-want-to-hear-message that mask work, but this question is more specific than that, i.e. wants to isolate the effect of source control as opposed to a group wearing masks.
And the Q also wants
answerers and voters to weight evidence about the spread of COVID-19 in real-life conditions above evidence from lab experiments
Ask yourselves if the other answers have even tried to follow these two requirements before you cast more stones at mine.
As of now I'm not aware of any RCTs that have been conducted on the issue, with Covid-19 as the infectious agent. And there has been plenty of debate, even among experts, whether it's mostly larger droplets or mostly aerosols that matter for this particular agent, so if you want a high degree of certainty, it's going to be difficult to rely on inferences from other/older studies on different agents...
Some such RCTs were conducted with influenza (and "influenza-like illnesses"), before this pandemic; at least one in China and a smaller one in France. The results were underwhelming, showing a trend but not reaching statistical significance (the Chinese study) or basically no effect whatsoever (France); their design may have been overly ambitious though as they measured transmission from confirmed index cases [randomized to with and without mask "treatment" groups] to other household members, with whom there were substantial opportunities for prolonged and varied forms of (unmonitored) interaction. The household "targets" were not required to wear any form protection.
A good number of sources argued that conducting this kind of trail (with subjects) during the pandemic is unethical. On the other hand, some Covid-19 challenge trials have been now (i.e. 2021) approved and conducted in the UK, but insofar nothing that involved testing masks as source-control, as far as I now. (The staff did wear PPE, as you might expect, and one could bet they were vaccinated too.)
So, we'll still have to rely on "natural experiments" outside of the technical laboratory ones that have tested mask efficacy in some artificial/proxy/model setting. (The other answers mostly cover these kinds of experiments.)
Of course, you could say I'm cherry picking, which is inevitable for this kind of natural experiments, but I'm going to point out two CDC case studies, one in which the infected sources wore masks (and seemingly the "targets" too), and one in which the source mostly did not, but the targets were reported to have mostly adhered to masks.
Among 139 clients exposed to two symptomatic hair stylists with confirmed COVID-19 while both the stylists and the clients wore face masks, no symptomatic secondary cases were reported; among 67 clients tested for SARS-CoV-2, all test results were negative. [...]
On May 12, 2020 (day 0), a hair stylist at salon A in Springfield, Missouri (stylist A), developed respiratory symptoms and continued working with clients until day 8, when the stylist received a positive test result for SARS-CoV-2, the virus that causes coronavirus disease 2019 (COVID-19). A second hair stylist (stylist B), who had been exposed to stylist A, developed respiratory symptoms on May 15, 2020 (day 3), and worked with clients at salon A until day 8 before seeking testing for SARS-CoV-2, which returned a positive result on day 10.
A total of 139 clients were directly serviced by stylists A and B from the time they developed symptoms until they took leave from work. Stylists A and B and the 139 clients followed the City of Springfield ordinance and salon A policy recommending the use of face coverings (i.e., surgical masks, N95 respirators, or cloth face coverings) for both stylists and clients during their interactions. Other stylists at salon A who worked closely with stylists A and B were identified, quarantined, and monitored daily for 14 days after their last exposure to stylists A or B. None of these stylists reported COVID-19 symptoms. [...]
Stylist A worked with clients for 8 days while symptomatic, as did stylist B for 5 days. During all interactions with clients at salon A, stylist A wore a double-layered cotton face covering, and stylist B wore a double-layered cotton face covering or a surgical mask. [...]
The Greene County Health Department (Missouri) conducted contact tracing for all 139 exposed clients back to the dates that stylists A and B first developed symptoms. The 139 clients were monitored after their last exposure at salon A. Clients were asked to self-quarantine for 14 days and were called or sent daily text messages to inquire about any symptoms; none reported signs or symptoms of COVID-19. Testing was offered to all clients 5 days after exposure, or as soon as possible for those exposed >5 days before contact tracing began. Overall, 67 (48.2%) clients volunteered to be tested, and 72 (51.8%) refused; all 67 nasopharyngeal swab specimens tested negative for SARS-CoV-2 by PCR. Telephone interviews were attempted 1 month after initial contact tracings to collect supplementary information. Among the 139 exposed clients, the Greene County Health Department interviewed 104 (74.8%) persons. [...]
Among the 104 interviewed clients, 102 (98.1%) reported wearing face coverings for their entire appointment, and two (1.9%) reported wearing face coverings part of the time (Table 2). Types of face covering used by clients varied; 49 (47.1%) wore cloth face coverings, 48 (46.1%) wore surgical masks, five (4.8%) wore N95 respirators, and two (1.9%) did not know what kind of face covering they wore. Overall, 101 (97.1%) interviewed clients reported that their stylist wore a face covering for the entire appointment; three did not know. When asked about the type of face coverings worn by the stylists, 64 (61.5%) reported that their stylist wore a cloth face covering (39; 37.5%) or surgical mask (25; 24.0%); 40 (38.5%) clients did not know or remember the type of face covering worn by stylists. When asked whether they had experienced respiratory symptoms in the 90 days preceding their appointment, 87 (83.7%) clients reported that they had not. Of those who did report previous symptoms, none reported testing for or diagnosis of COVID-19.
Six close contacts of stylists A and B outside of salon A were identified: four of stylist A and two of stylist B. All four of stylist A’s contacts later developed symptoms and had positive PCR test results for SARS-CoV-2. These contacts were stylist A’s cohabitating husband and her daughter, son-in-law, and their roommate, all of whom lived together in another household. None of stylist B’s contacts became symptomatic.
The teacher reported becoming symptomatic on May 19, but continued to work for 2 days before receiving a test on May 21. On occasion during this time, the teacher read aloud unmasked to the class despite school requirements to mask while indoors. [...]
The index patient [=teacher] became symptomatic on May 19 with nasal congestion and fatigue. This teacher reported attending social events during May 13–16 but did not report any known COVID-19 exposures and attributed symptoms to allergies. The teacher continued working during May 17–21, subsequently experiencing cough, subjective fever, and headache. The school required teachers and students to mask while indoors; interviews with parents of infected students suggested that students’ adherence to masking and distancing guidelines in line with CDC recommendations was high in class. [...]
During May 23–26, among the teacher’s 24 students, 22 students, all ineligible for vaccination because of age, received testing for SARS-CoV-2; 12 received positive test results. The attack rate in the two rows seated closest to the teacher’s desk was 80% (eight of 10) and was 28% (four of 14) in the three back rows (Fisher’s exact test; p = 0.036). During May 24–June 1, six of 18 students in a separate grade at the school, all also too young for vaccination, received positive SARS-CoV-2 test results. Eight additional cases were also identified, all in parents and siblings of students in these two grades.
All classrooms had portable high-efficiency particulate air filters and doors and windows were left open. [...]
Specimens for WGS were collected during May 26–June 12; all 18 positive specimens with detectable virus (cycle threshold value <32) were sequenced using ClearDx instruments (Clear Laboratories), Oxford Nanopore MinION sequencing technology, and SARS-CoV-2 ARTIC V3 protocol for amplicon sequencing. Consensus genome assembly was performed in Terra using Titan Clear Laboratories workflow. All sequences generated were classified as the Delta variant. A phylogenetic tree was constructed using the UShER pipeline and visualized using Auspice.us [...]. Eleven sequences were genetically indistinguishable from one another; seven sequences contained additional single nucleotide variations. Among the indistinguishable specimens, six were from students of the index patient, four were from students in the separate grade, and one was from a sibling of a student in the index patient’s class, suggesting that infections occurring in the two grades likely were part of the same outbreak. [...]
[Unfortunately,] the teacher’s specimen was unavailable for WGS, which prevented phylogenetic identification of the outbreak’s index patient. [...]
There is the obvious confounding factor that in the first natural experiment the virus variant involved was not delta, but in the latter one it was. Also, the total interaction time of the stylists with any of their clients was probably smaller than that of the teacher with the students... and the clients didn't interact with each other, but the students almost certainly did to some degree. So, yeah, this is hardly a prefect comparison.
Even in the first natural experiment, if you want to take the close/family contacts as controls... for one of the stylists they all developed infections, while for the other one, none of such "controls" developed infection. (It's worth noting though that only five (4.8%) of the clients wore professional grade PPE, i.e. N95 respirators, so the natural experiment in this regard was fairly different from hospital staff interaction with known infected persons.)
The 2nd natural experiment, i.e. the classroom one, was partially "ruined" by the fact that the teachers' sample was not subjected to deep/whole genome sequencing, so there's only confirmation that the children had the exact same or one-mutation-away [sub-]strain. (The tests were not conducted by the same lab.) Also, there isn't much in the CDC report on how long the teacher was unmasked; no structured interviewing of the students was reported (unlike for the clients of the stylists).
So, yeah, this is basically the kind of field evidence that the CDC has to base their recommendations on, besides the lab sneeze simulator experiments etc. from the other answers. The CDC itself says (or at least said in May 2021):
Data regarding the “real-world” effectiveness of community masking are limited to observational and epidemiological studies.
The first study they cite there is the one the two hair stylists, by the way. There's generally little or no effort made in that text to separate studies between source-control and wearer-protection when it comes to “real-world” effectiveness, and this is because it's hard to do that with observational data. Practically all evidence cited regarding source control (one para) is from technical studies on masks:
Source Control to Block Exhaled Virus
Multi-layer cloth masks block release of exhaled respiratory particles into the environment,[citations] along with the microorganisms these particles carry.[citations] Cloth masks not only effectively block most large droplets (i.e., 20-30 microns and larger)[citations] but they can also block the exhalation of fine droplets and particles (also often referred to as aerosols) smaller than 10 microns;[citations] which increase in number with the volume of speech[citations] and specific types of phonation.[citations] Multi-layer cloth masks can both block up to 50-70% of these fine droplets and particles[citations] and limit the forward spread of those that are not captured.[citations] Upwards of 80% blockage has been achieved in human experiments that have measured blocking of all respiratory droplets,[citations] with cloth masks in some studies performing on par with surgical masks as barriers for source control.[citations]
They do cite something like 12 different papers on such technical/lab studies on masks; omitting the list here since the page is open-access. A typical paper cited e.g. for the last sentence is Kawaoka et al..
Here, we developed an airborne transmission simulator of infectious SARS-CoV-2-containing droplets/aerosols produced by human respiration and coughs and assessed the transmissibility of the infectious droplets/aerosols and the ability of various types of face masks to block the transmission. We found that cotton masks, surgical masks, and N95 masks all have a protective effect with respect to the transmission of infective droplets/aerosols of SARS-CoV-2 and that the protective efficiency was higher when masks were worn by a virus spreader.
The keyword there is "simulator" as far as “real-world” effectiveness is concerned. (The simulator was basically two mannequin heads [facing each other] enclosed in an airtight box, with some pumping gear to simulate breathing through the heads.) Also note the study's conclusion that masks had more effect when worn by the source; it's probably from studies like these that such conclusion made its way into the guidelines.
The Cheng et al. paper that has been laconically quoted in another answer actually has something to say (based on its theoretical model/simulation) about source-control too. This is most detailed in its supplementary material.
Figure 3 shows that source control is more effective than wearer protection in reducing Pinf,pop [probability of infection, at population level]
via airborne transmission of SARS-CoV-2. Besides airborne transmission, other pathways have
also been suggested for the transmission of respiratory viruses like SARS-CoV-2. For example,
through direct or indirect contact between people and contaminated surfaces; or through
respiratory droplets that are larger than 100 μm and would typically fall to the ground in seconds
within 2 m of the source and not affect distant people (30, 49). Source control is very efficient in
removing large droplets >100 μm (~100% efficiency) as illustrated in Figs. 4 and S5. Thus,
source control can also reduce eye infections by droplets, which would not be prevented by
wearer protection. Moreover, source control reduces the overall release of respiratory viruses,
and thus their availability for contact transmission.
So the theory/simulation in this paper (which is largely a model/simulation of a large hall: 500m2with 200 people -- based on a Fangcang hospital) seems to agree with the physical simulation data from previous "micro" works that used only two "people". I'm mentioning it as being towards some "expert/model consensus" on source control being the more significant mechanism by which masks work.