Related
We have recently taken over a project from an outsourcing company. This project uses Moles and Pex for unit testing, but since we have not had the project for long, I am not very familiar with the frameworks.
That being said, we are busy upgrading this project to run in .Net 4. I have resolved most of the issues that have jumped out, but there is one that I cannot get a handle on. Some of the unit tests cannot compile because of the error:
Could not load file or assembly 'Example.Assembly, Version=0.0.0.0,
Culture=neutral, PublicKeyToken=null' or one of its dependencies. The
system cannot find the file specified.
The part that baffles me is that it is a project reference and the assembly is being copied to the output directory of the unit test. Most of the other project references are found and I cannot spot any difference between the ones that work and the ones that do not. I am not sure if this problem has to do with the pex/moles frameworks, but I thought I would mention it.
I have tried the usual things of removing and adding all the references and regenerating the moles assemblies.
Has anyone else run into this problem? Any help would be greatly appreciated.
EDIT1: Ok, after some more investigation into the build output, it appears as if it is not moles, but the .accessor files that are not generated correctly. I get the exact same problem as asked in Unit test project cannot find assembly under test (or dependencies), but unlike his problem, mine does not go away after deleting the accessor.
EDIT2: Turns out is is a program called Publicize.exe which falls over with that error. Still no idea why though. Looking at Fusion logs is looks like it does not search under the working directory for the dll that it is trying to generate the accessors for. Running it manually on a bunch of assemblies from our solution, I find it works on some, but not on others. I have not been able to identify a difference between the ones that work and the ones that don't, though.
Thanks
Ah, yes. I have read this story many times, and have the tee shirt. I run through my usual Moles first-aid kit, when encountering any issue, including this one.
Perhaps, this question will provide some help: Am I the only one getting "Assembly Not Available in the Currently Targeted Framework"?
Ensure the Moles framework is properly installed on the workstation and/or build server
Ensure the Moles assemblies are being built (see the excluded "Moles Assemblies" directory)
Check your build profile -- it may need to be set to full framework profile
Triple check your output destinations and post-build commands -- I have seems some solutions that copy the output to another location
Try using the Visual Studio Pex/Moles extension, if you are not already doing so
An invasive fix-all process is to simply create an all-new solution, projects, and test projects, and then copy the existing code files into them. It's surprising how many issues can be resolved for various project-related errors. Basically, a hard reboot for the entire solution.
Since you are updating to .NET 4, you may as well go to 4.5, and used the productized version of Moles, called "Fakes". You'll find Fakes in the Visual Studio 2012 release candidate. This significant feature hasn't received much attention.
Currently I have 4 solutions that are independent projects, however there is quite a bit of duplicated code when it comes to a few areas of the applications.
At the moment it is simply a few forms and their associated code.
When I want to change or improve something, I have to copy and paste to all relevant projects.
I looked at creating a new project within one of the solutions for the .dll/class library, but I felt that this was incorrect. (Please say if I am wrong).
As it is a component for all the applications, I decided to create a new solution for the .dll/class library and am looking at moving the shared code over to that - but, having never gone down this route before, what are my options from here?
Am I able to then include this solution within the others if I need to make a simple change and have it updated in all the projects or instead, should I always be working on the shared component in a separate instance of Visual Studio, outside of the applications using it?
That's exactly the right way to handle this situation.
You can include projects in multiple solutions by right-clicking the solution and selecting Add Existing Project...
Any changes you then make will appear in all solutions. The only problem this leads to is that it's possible to break one solution from another. This is where automated builds on commit to source control come into their own.
Put shared codes in separate Solution/Project as Class Library,
In post build event of shared projects copy dll's to a specific directory,
Add shared dll's from this directory to other projects/solutions
By doing this each time you build your consumer projects, they will use latest dll's automatically.
Moving the common code into a separate shared assembly is an excellent option.
One thing to think about is to keep your common business logic or business object type code separate from UI related code like custom controls - if you need to then have two common assemblies. This is more work initially, but makes things way easier further down the track when you need to make UI changes or change the control suite you are using.
Note: This is for a shop that works in C++, C++/CLI, and C# with some products being delivered as a combination of all three.
We currently have a rule that a project should have one and only one containing solution. The rule was originally instated because the source control plug-in for Visual Studio was unable to cope with projects that were contained in multiple solutions, always attempting to change the source control bindings when changing from one solution to another.
For other reasons, we are going to stop using the source control plug-in altogether (not dropping source control, just the brain-dead plug-in). It re-opens the question of whether or not to continue the policy of restricting projects to be contained in only one solution.
We have quite a bit of code in libraries, dlls, and assemblies that are consumed by multiple deliverable products, and we currently control this with a system of inter-solution dependency management whereby if one is working in the solution for an end-product, it is a simple matter to request a build of the dependency solutions, which kicks off other instances of Visual Studio to build them. The system works, but has the following drawbacks:
The solutions for the common projects are often too fat, containing a few projects needed by the product being currently developed and usually a lot more that aren't needed for the development work at hand. They have merely been lumped into a sort-of catch-all solution that covers projects that are merely loosely related. This results in extended build times due to the compilation of unneeded code.
Where we have tried to address the above fat solutions, we are often left with a too skinny solution that contains just one project. Each of these requires additional configuration in the inter-solution dependency management system. This, too, can increase build times as multiple instances of Visual Studio are spun up.
I am considering revising the policy to allow a project used in multiple deliverable products to be contained in more than one solution. We could eliminate the inter-solution dependency management and severely reduce the number of solutions (down to one per product). I am concerned about the amount of work this reorganization will take and whether or not it will be worth the effort. I'm afraid I won't even be able to detect the potential benefits until the team has been working with it for a while. I also foresee a couple of potential issues which are the true questions here.
Do a large number of projects in a single solution pose any stability problems in Visual Studio 2005 (our current development platform)? Does it get any better in VS 2008 or VS 2010?
If a project is contained in more than one solution, how can one manage the effects on the multiple solutions should project configurations be modified (such as adding a new configuration)?
To anyone already working in an environment with one solution per deliverable product, with common components as projects contained in multiple solutions: Have you encountered any significant drawbacks to such a configuration?
No, use as many solutions as is convenient.
For example, we have a large solution that builds about 53 projects. It includes an installer and uninstaller so that they can be conveniently built by our build server, but if I want to work on those applications, I load them in their own solutions (1 project each), rather than wasitng all that time loading 69 other unnecessary projects. They don't have any dependencies on the other projects, so there's no problem at all doing it this way (indeed, it's a lot more efficient as the solution loads and builds way faster)
The only caveat of multiple solutions is that you have to be careful in any case where you don't build a dependency (e.g. if you don't build a library, then you need to be careful that it gets built whenever the source code for it changes, because it will no longer be built automatically for you)
edit
To answer the last part of your question (because SO takes the question away while you're editing the answer!), the only problem with Solutions in VS2005 is that it is very slow with a lot of projects in them. VS2008 is much faster at loading a large solution, but the main hit is just that the more projects there are in a solution, the longer it takes to build (even if it is just skipping projects that don't need to be built, it takes a long time)
If you add a new solution configuration, then just make one uber-solution (or just keep your existing one!) and then make your changes in that so that they are applied to all porjects in one go. You will still have to add that new configuration to any separate solutions you want to use, but if they're one-project solutions you can just delete them, load up the .csproj and it'll be rebuilt automatically with all the configs in it. And adding new configurations probably isn't going to happen very often.
The main problems with having one project in several solutions are:
If you make a change in the project, it may cause a compile error in another solution in which it is used.
The solutions must constantly be updated to keep up with the changes in the common projects
The benefit of having a project in each solution that it is used is:
It is easier to debug when you can follow the problem done to the souce code.
Another way to do it would be to have each solution use the the dll's of the common projects that they require. Then each solution can decide when they want to change which version they use.
The number of projects in a solution is not a major problem. But it will make the sollution slower to load and build. We have solutions of over 50 projects.
I too work in an environment that has many common projects that are used in several deliverables. Our policy has been one solution per deliverable application. So a common project may be included in several solutions. This has worked well for us. The common projects, while contained in separate solutions, all map to the same source code repository locations. So if the common code is changed, all containing applications are affected. Our continuous integration system serves as a check that other applications do not break as a result of the change to the common code.
The more projects you have in a VS solution, the longer it takes to load and it just sounds complicated having to manage different build configurations and dependencies in a single solution.
We have a "master solution" that contains over 100 projects. This took forever to load in VS.2005 until Microsoft released Intellisense fixes described here.
Our automated "continuous integration" build procedure always builds the master solution. For development, I divide the "master solution" into smaller solutions that contain a subset of related projects. Maintaining this structure takes some time when new projects are added, but is well worth it.
The only problem I've discovered with projects in multiple solutions is avoided by using $(ProjectDir) instead of $(SolutionDir) in build rules, as the latter depends on how the project is loaded.
Tools for SLN file (Visual Studio Solution file) contains a tool that:
Make it possible to create filters for a SLN file. The way it work is
that when the filter file is opened, a temporary solution is created
dynamically. The created solution file contain only the projects that
are specified in the filter, and all the dependencies of those
projects. This make it possible to have a mega-solution that contain
all the projects needed to build the complete system but still have
the possibility to work efficiently on a subset of the system.
This removes a big part of the cost of maintain sub-set solution files for speed of development.
However you need to ask youself if you have more project files then you need, as combining project files can speed up compiles and loading of soltuions greatly.
There is now a free Visual Studio extension, "Funnel", that solves this problem by loading predefined subsets of the solution. The link is for VS2012-VS2015; there is also a VS2010 version referenced on the extension home page.
You can set up multiple profiles (i.e. project combinations). These are stored in a parallel solution.sln.vsext file and that file can be committed and shared with your team. While VS2015 has gotten much better at handling large, 100+ C/C++/C# project solutions, I find it particularly useful for excluding makefile projects that are built every time the solution is built.
I have a C# project which includes one exe and 11 library files. The exe references all the libraries, and lib1 may reference lib2, lib3, lib4, etc.
If I make a change to a class in lib1 and built the solution, I assumed that only lib1 and the exe would need to be changed. However, all dll's and the exe are being built if I want to run the solution.
Is there a way that I can stop the dependencies from being built if they have not been changed?
Is the key this phrase? "However, all dll's and the exe are being built if I want to run the solution"
Visual Studio will always try to build everything when you run a single project, even if that project doesn't depend on everything. This choice can be changed, however. Go to Tools|Options|Projects and Solutions|Build and Run and check the box "Only build startup projects and dependencies on Run". Then when you hit F5, VS will only build your startup project and the DLLs it depends on.
I just "fixed" the same problem with my VS project. Visual Studio did always a rebuild, even if didn't change anything. My Solution: One cs-File had a future timestamp (Year 2015, this was my fault). I opened the file, saved it and my problem was solved!!!
I am not sure if there is a way to avoid dependencies from being built. You can find some info here like setting copylocal to false and putting the dlls in a common directory.
Optimizing Visual Studio solution build - where to put DLL files?
We had a similar problem at work. In post-build events we were manually embedding manifests into the outputs in the bin directory. Visual Studio was copying project references from the obj dir (which weren't modified). The timestamp difference triggered unnecessary rebuilds.
If your post-build events modify project outputs then either modify the outputs in the bin and obj dir OR copy the modified outputs in the bin dir on top of those in the obj dir.
You can uncheck the build option for specified projects in your Solution configuration:
(source: microsoft.com)
You can can create your own solution configurations to build specific project configurations...
(source: microsoft.com)
We actually had this problem on my current project, in our scenario even running unit tests (without any code changes) was causing a recompile. Check your build configuration's "Platform".
If you are using "Any CPU" then for some reason it rebuilds all projects regardless of changes. Try using processor specific builds, i.e. x86 or x64 (use the platform which is specific to the machine architecture of your machine). Worked for us for x86 builds.
(source: episerver.com)
Now, after I say this, some propeller-head is going to come along and contradict me, but there is no way to do what you want to do from Visual Studio. There is a way of doing it outside of VS, but first, I have a question:
Why on earth would you want to do this? Maybe you're trying to save CPU cycles, or save compile time, but if you do what you're suggesting you will suddenly find yourself in a marvelous position to shoot yourself in the foot. If you have a library 1 that depends upon library 2, and only library 2 changes, you may think you're OK to only build the changed library, but one of these days you are going to make a change to library 2 that will break library 1, and without a build of library 2 you will not catch it in the compilation. So in my humble opinion, DON'T DO IT.
The reason this won't work in VS2005 and 2008 is because VS uses MSBuild. MSBuild runs against project files, and it will examine the project's references and build all referenced projects first, if their source has changed, before building the target project. You can test this yourself by running MSBuild from the command line against one project that has not changed but with a referenced project that has changed. Example:
msbuild ClassLibrary4.csproj
where ClassLibrary4 has not changed, but it references ClassLibrary5, which has changed. MSBuild will build lib 5 first, before it builds 4, even though you didn't mention 5.
The only way to get around all these failsafes is to use the compiler directly instead of going through MSBuild. Ugly, ugly, but that's it. You will basically be reduced to re-implementing MSBuild in some form in order to do what you want to do.
It isn't worth it.
Check out the following site for more detailed information on when a project is built as well as the differences between build and rebuild.
I had this problem too, and noticed these warning messages when building on Windows 7 x64, VS2008 SP1:
cl : Command line warning D9038 : /ZI is not supported on this platform; enabling /Zi instead
cl : Command line warning D9007 : '/Gm' requires '/Zi'; option ignored
I changed my project properties to:
C/C++ -> General -> Debug Information Format = /Zi
C/C++ -> Code Generation -> Enable Minimal Build = No
After rebuilding I switched them both back and dependencies work fine again. But prior to that no amount of cleaning, rebuilding, or completely deleting the output directory would fix it.
I don't think there's away for you to do it out of the box in VS. You need this add-in
http://workspacewhiz.com/
It's not free but you can evaluate it before you buy.
Yes, exclude the non-changing bits from the solution. I say this with a caveat, as you can compile in a way where a change in build number for the changed lib can cause the non built pieces to break. This should not be the case, as long as you do not break interface, but it is quite common because most devs do not understand interface in the .NET world. It comes from not having to write IDL. :-)
As for X projcts in a solution, NO, you can't stop them from building, as the system sees a dependency has changed.
BTW, you should look at your project and figure out why your UI project (assume it is UI) references the same library as everything else. A good Dependency Model will show the class(es) that should be broken out as data objects or domain objects (I have made an assumption that the common dependency is some sort of data object or domain object, of course, but that is quite common). If the common dependency is not a domain/data object, then I would rethink my architecture in most cases. In general, you should be able to create a path from UI to data without common dependencies other than non-behavioral objects.
Not sure of an awesome way to handle this, but in the past if I had a project or two that kept getting rebuilt, and assuming I wouldn't be working in them, I would turn the build process off for them.
Right click on the sln, select configuration manager and uncheck the check boxes. Not perfect, but works when Visual Studio isn't behaving.
If you continue to experience this problem, it may be due to a missing or out of date calculated dependency (like a header) that is listed in your project, but does not exist.
This happens to me especially common after migrating to a new version (for example: from 2012 to 2013) because VS may have recalculated dependencies in the conversion, or you are migrating to a new location.
A quick check is to double-click every file in offending project from solution explorer. If you discover a file does not exist, that is your problem.
Failing a simple missing file: You may have a more complicated build date relationship between source and target. You can use a utility to find out what front-end test is triggering the build. To get that information you can enable verbose CPS logging. See: Andrew Arnott - Enable C++ and Javascript project system tracing (http://blogs.msdn.com/b/vsproject/archive/2009/07/21/enable-c-project-system-logging.aspx). I use the DebugView option. Invaluable tool when you need it.
(this is a C# specific question, but a different post was merged as identical)
UPDATE:
This is one of my most-visited questions, and yet I still haven't really found a satisfactory solution for my project. One idea I read in an answer to another question is to create a tool which can build solutions 'on the fly' for projects that you pick from a list. I have yet to try that though.
How do you structure a very large application?
Multiple smallish projects/assemblies in one big solution?
A few big projects?
One solution per project?
And how do you manage dependencies in the case where you don't have one solution.
Note: I'm looking for advice based on experience, not answers you found on Google (I can do that myself).
I'm currently working on an application which has upward of 80 dlls, each in its own solution. Managing the dependencies is almost a full time job. There is a custom in-house 'source control' with added functionality for copying dependency dlls all over the place. Seems like a sub-optimum solution to me, but is there a better way? Working on a solution with 80 projects would be pretty rough in practice, I fear.
(Context: winforms, not web)
EDIT: (If you think this is a different question, leave me a comment)
It seems to me that there are interdependencies between:
Project/Solution structure for an application
Folder/File structure
Branch structure for source control (if you use branching)
But I have great difficulty separating these out to consider them individually, if that is even possible.
I have asked another related question here.
Source Control
We have 20 or 30 projects being built into 4 or 5 discrete solutions. We are using Subversion for SCM.
1) We have one tree in SVN containing all the projects organised logically by namespace and project name. There is a .sln at the root that will build them all, but that is not a requirement.
2) For each actual solution we have a new trunks folder in SVN with SVN:External references to all the required projects so that they get updated from their locations under the main tree.
3) In each solution is the .sln file plus a few other required files, plus any code that is unique to that solution and not shared across solutions.
Having many smaller projects is a bit of a pain at times (for example the TortoiseSVN update messages get messy with all those external links) but does have the huge advantage that dependancies are not allowed to be circular, so our UI projects depend on the BO projects but the BO projects cannot reference the UI (and nor should they!).
Architecture
We have completely switched over to using MS SCSF and CAB enterprise pattern to manage the way our various projects combine and interact in a Win Forms interface. I am unsure if you have the same problems (multiple modules need to share space in a common forms environment) but if you do then this may well bring some sanity and convention to how you architect and assemble your solutions.
I mention that because SCSF tends to merge BO and UI type functions into the same module, whereas previously we maintained a strict 3 level policy:
FW - Framework code. Code whose function relates to software concerns.
BO - Business Objects. Code whose function relates to problem domain concerns.
UI - Code which relates to the UI.
In that scenario dependancies are strictly UI -> BO -> FW
We have found that we can maintain that structure even while using SCSF generated modules so all is good in the world :-)
To manage dependencies, whatever the number of assemblies/namespaces/projects you have, you can have a glance at the tool NDepend.
Personnaly, I foster few large projects, within one or several solutions if needed. I wrote about my motivations to do so here: Benefit from the C# and VB.NET compilers perf
I think it's quite important that you have a solution that contains all your 80 projects, even if most developers use other solutions most of the time. In my experience, I tend to work with one large solution, but to avoid the pain of rebuilding all the projects each time I hit F5, I go to Solution Explorer, right-click on the projects I'm not interested in right now, and do "Unload Project". That way, the project stays in the solution but it doesn't cost me anything.
Having said that, 80 is a large number. Depending on how well those 80 break down into dicrete subsystems, I might also create other solution files that each contain a meaningful subset. That would save me the effort of lots of right-click/Unload operations. Nevertheless, the fact that you'd have one big solution means there's always a definitive view of their inter-dependencies.
In all the source control systems that I've worked with, their VS integration chooses to put the .sln file in source control, and many don't work properly unless that .sln file is in source control. I find that intriguing, since the .sln file used to be considered a personal thing, rather than a project-wide thing. I think the only kind of .sln file that definitely merits source control is the "one-big-solution" that contains all projects. You can use it for automated builds, for example. As I said, individuals might create their own solutions for convenience, and I'm not against those going into source control, but they're more meaningful to individuals than to the project.
I think the best solution is to break it in to smaller solutions. At the company I currently work for, we have the same problem; 80 projects++ in on solution. What we have done, is to split into several smaller solutions with projects belonging together. Dependent dll's from other projects are built and linked in to the project and checked in to the source control system together with the project. It uses more disk space, but disk is cheap. Doing it this way, we can stay with version 1 of a project until upgrading to version 1.5 is absolutely necessary. You still have the job with adding dll's when deciding to upgrade to a other version of the dll though. There is a project on google code called TreeFrog that shows how to structure the solution and development tree. It doesn't contain mush documentation yet, but I guess you can get a idea of how to do it by looking at the structure.
A method that i've seen work well is having one big solution which contains all the projects, for allowing a project wide build to be tested (No one really used this to build on though as it was too big.), and then having smaller projects for developers to use which had various related projects grouped together.
These did have depencies on other projects but, unless the interfaces changed, or they needed to update the version of the dll they were using, they could continue to use the smaller projects without worrying about everything else.
Thus they could check-in projects while they were working on them, and then pin them (after changing the version number), when other users should start using them.
Finally once or twice a week or even more frequently the entire solution was rebuild using pinned code only, thus checking if the integration was working correctly, and giving testers a good build to test against.
We often found that huge sections of code didn't change frequently, so it was pointless loading it all the time. (When you're working on the smaller projects.)
Another advantage of using this approach is in certain cases we had pieces of functionality which took months to complete, by using the above approach meant this could continue without interrupting other streams of work.
I guess one key criteria for this is not having lots of cross dependencies all over your solutions, if you do, this approach might not be appropriate, if however the dependencies are more limited, then this might be the way to go.
For a couple of systems I've worked on we had different solutions for different components. Each solution had a common Output folder (with Debug and Release sub-folders)
We used project references within a solution and file references between them. Each project used Reference Paths to locate the assemblies from other solutions. We had to manually edit the .csproj.user files to add a $(Configuration) msbuild variable to the reference paths as VS insists on validating the path.
For builds outside of VS I've written msbuild scripts that recursively identify project dependencies, fetch them from subversion and build them.
I gave up on project references (although your macros sound wonderful) for the following reasons:
It wasn't easy to switch between different solutions where sometimes dependency projects existed and sometimes didn't.
Needed to be able to open the project by itself and build it, and deploy it independently from other projects. If built with project references, this sometimes caused issues with deployment, because a project reference caused it to look for a specific version or higher, or something like that. It limited the mix and match ability to swap in and out different versions of dependencies.
Also, I had projects pointing to different .NET Framework versions, and so a true project reference wasn't always happening anyways.
(FYI, everything I have done is for VB.NET, so not sure if any subtle difference in behavior for C#)
So, I:
I build against any project that is open in the solution, and those that aren't, from a global folder, like C:\GlobalAssemblies
My continuous integration server keeps this up to date on a network share, and I have a batch file to sync anything new to my local folder.
I have another local folder like C:\GlobalAssembliesDebug where each project has a post build step that copies its bin folder's contents to this debug folder, only when in DEBUG mode.
Each project has these two global folders added to their reference paths. (First the C:\GlobalAssembliesDebug, and then C:\GlobalAssemblies). I have to manually add this reference paths to the .vbproj files, because Visual Studio's UI addes them to the .vbprojuser file instead.
I have a pre-build step that, if in RELEASE mode, deletes the contents from C:\GlobalAssembliesDebug.
In any project that is the host project, if there are non dlls that I need to copy (text files outputted to other project's bin folders that I need), then I put a prebuild step on that project to copy them into the host project.
I have to manually specify the project dependencies in the solution properties, to get them to build in the correct order.
So, what this does is:
Allows me to use projects in any solution without messing around with project references.
Visual Studio still lets me step into dependency projects that are open in the solution.
In DEBUG mode, it builds against open loaded projects. So, first it looks to the C:\GlobalAssembliesDebug, then if not there, to C:\GlobalAssemblies
In RELEASE mode, since it deletes everything from C:\GlobalAssembliesDebug, it only looks to C:\GlobalAssemblies. The reason I want this is so that released builds aren't built against anything that was temporarily changed in my solution.
It is easy to load and unload projects without much effort.
Of course, it isn't perfect. The debugging experience is not as nice as a project reference. (Can't do things like "go to definition" and have it work right), and some other little quirky things.
Anyways, that's where I am on my attempt to make things work for the best for us.
We have one gigantic solution on the source control, on the main branch.
But, every developer/team working on the smaller part of the project, has its own branch which contains one solution with only few projects which are needed. In that way, that solution is small enough to be easily maintenaced, and do not influence on the other projects/dlls in the larger solution.
However, there is one condition for this: there shouldn't be too much interconnected projects within solution.
OK, having digested this information, and also answers to this question about project references, I'm currently working with this configuration, which seems to 'work for me':
One big solution, containing the application project and all the dependency assembly projects
I've kept all project references, with some extra tweaking of manual dependencies (right click on project) for some dynamically instantiated assemblies.
I've got three Solution folders (_Working, Synchronised and Xternal) - given that my source control isn't integrated with VS (sob), this allows me to quickly drag and drop projects between _Working and Synchronised so I don't lose track of changes. The XTernal folder is for assemblies that 'belong' to colleagues.
I've created myself a 'WorkingSetOnly' configuration (last option in Debug/Release drop-down), which allows me to limit the projects which are rebuilt on F5/F6.
As far as disk is concerned, I have all my projects folders in just one of a few folders (so just one level of categorisation above projects)
All projects build (dll, pdb & xml) to the same output folder, and have the same folder as a reference path. (And all references are set to Don't copy) - this leaves me the choice of dropping a project from my solution and easily switching to file reference (I've got a macro for that).
At the same level as my 'Projects' folder, I have a 'Solutions' folder, where I maintain individual solutions for some assemblies - together with Test code (for example) and documentation/design etc specific to the assembly.
This configuration seems to be working ok for me at the moment, but the big test will be trying to sell it to my colleagues, and seeing if it will fly as a team setup.
Currently unresolved drawbacks:
I still have a problem with the individual assembly solutions, as I don't always want to include all the dependent projects. This creates a conflict with the 'master' solution. I've worked around this with (again) a macro which converts broken project references to file references, and restores file references to project references if the project is added back.
There's unfortunately no way (that I've found so far) of linking Build Configuration to Solution Folders - it would be useful to be able to say 'build everything in this folder' - as it stands, I have to update this by hand (painful, and easy to forget). (You can right click on a Solution Folder to build, but that doesn't handle the F5 scenario)
There is a (minor) bug in the Solution folder implementation which means that when you re-open a solution, the projects are shown in the order they were added, and not in alphabetical order. (I've opened a bug with MS, apparently now corrected, but I guess for VS2010)
I had to uninstall the CodeRushXPress add-in, because it was choking on all that code, but this was before having modified the build config, so I'm going to give it another try.
Summary - things I didn't know before asking this question which have proved useful:
Use of solution folders to organise solutions without messing with disk
Creation of build configurations to exclude some projects
Being able to manually define dependencies between projects, even if they are using file references
This is my most popular question, so I hope this answer helps readers. I'm still very interested in further feedback from other users.