...since that is what it meant to be.
To allow this chance, I've now added a default ctor to lib::Literal,
defaulting to the Symbol::EMPTY (the interned empty string)
The class Literal is used as a thin wrapper to mark the fact that
some string parameter or value is assumed to be given *literally*
For the contract this indicates
- that storage is somewhere
- storage is not owned and managed by Literal
- yet storage guaranteed to exist during the whole lifetime of the program
- Literal can not be altered
- Literal is transparently convertible to const char *
Currently I am in the course of building some path abstraction, and for that
task it makes sense to hold an array of Literals (instead of pointers), just
because it expresses the intent way more clear. I do not see anything in the
above mentioned contract to prohibit a default constructed Literal, with the
empty string being the most obvious choice.
Note: there is the class Symbol, which derives from Literal. Symbol takes
arbitrary strings, but *interns* them into a static symbol table.
...under the assumption that the content is normalised,
which means
- leading NULL is changed to Symbol::EMPTY
- missing elements in the middle are marked as "*"
- trailing NULL in extension storage is handled by adjusting nominal extension size
after various fruitless attempts to rely somehow on the array variant of unique_ptr,
I ended up with a hand coded version of an heap allocated array, managed automatically
as it turned out, the solution from yesterday works only with uniform argument lists,
but not with arbitrarily mixed types. Moreover the whole trickery with the
indices was shitty -- better use a predicate decision on template argument level.
This simple solution somehow just didn't occur to me...
...still somewhat unsatisfactory, because
- no clear compile error message when invoking pickArg with insufficient arguments
- the default initialisation case in SelectVararg is duplicated and messy
some time ago we abandoned our own tuple type in favour of std::tuple
Since then, the helpers and ported utilities provide some generic helpers
to deal with variadic argument sequences, especially to build index sequences,
which in turn can be used to "pick" individual arguments from a variadic parameter pack.
The expectation is for this part of the support library gradually to grow and
in parts to replace the existing type sequence processing helpers. The expectation
is that we'll retain the basic type sequence, lib::meta::Types, but retrofit it
to rely on variadic arguments
since the adoption of C++11, we gradually transition our metaprogramming helpers
to support and rely on variadic template parameters. For the time being,
we just augment existing facilities when it comes in handy, yet some more
heavyweight lifting and overall clean-up remains to be done eventually.
we allow assignment to the element embedded within the wrapper.
Yet obviously we need specific implementations for assignment
to the container itself. Thus we define the templated
assignment operator such as to render the explicit specialisation
a better match than anything generated from the templated
operator
basically DiffMessage has a "take everything" ctor, which happens
to match on type DiffMessage itslef, since the latter is obviously
a Lumiera Forward Operator. Unfortunately the compiler now considers
this "take everyting" ctor as copy constructor. Worse even, such a
template generated ctor qualifies as "best match".
The result was, when just returing a DiffMessage by value form a
function, this erroneous "copy" operation was invoked, thus wrapping
the existing implementation into a WrappedLumieraIterator.
The only tangible symptom of this unwanted storage bloat was the fact
that our already materialised diagnostics where seemingly "gone". Indee
they weren't gone for real, just covered up under yet another layer of
DiffMessage wrapping another Lumiera Forward Iterator
by moving, we can avoid the generation of up to 3 additional shared copies
of the DataHandle. The whole invocation now works without touching any shared count
and thus without incurring a memory barrier...
this becomes more relevant now, since the actual MutationMessage iterators
are implemented in terms of a shared_ptr to IterSource. Thus, when building
processing pipelines, we most definitively want to move that smart-ptr into
the destination, since this avoids touching the shared count and thus avoids
generating unnecessary memory barriers.
...allows us to get rid of quite some boost-includes
Incidentally, "our own" implementation is equivalent to both the
boost implementation and the implementation from C++14
It is just a bit more concise to write.
since we do not want to increase the footprint, we're bound to reuse
an existing VTable -- so IterAdapter itself is our only option.
Unfortunately we'll need to pass that through one additional
decoration layer, which is here the iterator; to be able to
add our string conversion there, we need to turn that into
a derived class and add a call to access the underlying
container, which gets us into element type definition mess....
now this highlights the unsettled decision still the more,
as can be seen by all that unnecessary copying. Basically we move the
Diff into the lambda-closure, from there into an anonymous instance,
from there into the embedded Buffer in MutationMessage, which again
just happens to sit in the closure storage when the action is invoked.
And all of this copying just to move the DiffMessage for consumption
into the TreeMutator...
thus by #1066 we should really get rid of the MutationMessage class altogether!
actually I do not know much regarding the actual situation when,
within the Builder run, we're able to detect a change and generate
a diff description. However, as a first step, I'll pick IterSrouce
as a base interface and use a "generation context", which is to be
passed by shared-ptr
again surprising how such fundamental bugs can hide for years...
Here the reason is that IterAdapter leaves the representation of "NIL" to
its instantiation / users; some users (here in for example the ScopedCollection)
can choose to allow for different representations of "NIL", but the comparison
provided by IterAdapter just compares the embedded pos by face value.
seems like most usages will want to expose this kind of diagnostics for unit testing
and in fact the queue or stack nature is the primary nature of this entity,
while iterability comes as additional trait
- concept for a first preliminary implementation of dispatch into the UI thread
- define an integration effort to build a complete working communication chain
...again to make it work with GCC-5,
also to allow more leeway using various compilers
Explanation: we use a helper function to abbreviate the
demangled type names to make diagnostic ouput more readable.
Obviously such a function needs to be adjusted to the
way concrete compilers generate their type output; GCC-5
slightly differs to GCC-4.9 here, so I've made the regular
expressions a bit more flexible
we have a catch-all template operator to get a string converted
or pretty printed output from "any object". Unfortunately
this overload counts equivalent to another overload by
the IO manipulators. Solution is to define both operarators
similar in the first argument, thus turing the overload
for the IO manipulators into the more specific overload
due to the explicitly given second argument
this seems like an obvious functionality and basically harmless,
since commands are designed to be inherently stateful, which is reflected
in all the internal storage holders to expos an assignment operator
(even while the actual implementation is based on placement new instead
of assigning values into the storage, and thus even supports immutable
values). The only possible ramification is that argument values must
be default constructible
in accordance to the design changes concluded yesterday.
- in the standard cases we now check the global registry first
- automatically create anonymous clone copy from global commands
- reorganise code internally to use common tail implementation
this might turn into lock contention problem, but better optimise
a correct implementation than fix a fast yet broken one.
Hint: SessionCommandFunction_test demonstrates that the
symbol table can be corrupted by creating Symbol instances
in parallel without proper locking. So yes, this is for real.
since Symbol instance are now backed by a symbol table,
we can use a much faster hash function by just hashing the
pointer into the symbol table, since the Symbol string content
is already checked at initialisation.
Up to now, we tolerated null pointers in Literal instances.
But we can not tolerate passing a null cString to Symbol initialisation.
Rather, hereby we introduce a dedicated "bottom" Symbol, a valid "null object"
For this task, I've also investigated to use boost::operators
This would only incur a negligible penalty on build times and executable sizes,
however, I don't consider the boost based solution to improve readability,
since many of these comparisons are tricky or subtly different.
Moreover, since boost::operators needs to be mixed-in, the initialisation
of Symbol objects becomes difficult, not to mention the additional base class
information visible in the debugger when inspecting Symbol or Literal objects
For that reason, I decided *against* using Boost here and coded up
all the operators in all combinations manually
...which means, from now on identical input strings
will produce the same Symbol object (embedded pointer).
TODO: does not handle null pointers passed in as c-String properly
obsoleted by C++11
* in most cases, it can be replaced by an explicit conversion operator
* especially for the Lumiera Forward Iterators, we need an implicit conversion
This changeset fixes a huge pile of problems, as indicated in the
error log of the Doxygen run after merging all the recent Doxygen improvements
unfortunately, auto-linking does still not work at various places.
There is no clear indication what might be the problem.
Possibly the rather unstable Sqlite support in this Doxygen version
is the cause. Anyway, needs to be investigated further.
it is not *that* hard to behave in a somewhat sane manner here.
And even more: this *is* basically the symbol table implementation we need.
Thus we only need to build the right front-end now...
...otherwise our log will be flooded with command definition messages soon
NOTE: to see all command definitions happening, set into environment:
NOBUG_LOG='command:TRACE
...since there is not any test coverage for this trait, which
turned out to be quite deeply rooted in the system by now and
handles several rather subtle special cases
...and move the tail-call of the template instantiation into try.cpp
This experiment clearly shows the discrepancy now:
- binding a member pointer directly into a function object will expand the argument list
- but binding a similar lambda into a function object won't
(it is not necessary due to the context capture)
The result is that we need to drop support for one of those cases,
and it is clear that the member poiter will be the looser...
As a first step towards a gradual rework of our function metaprogramming helpers,
this change prepends a generic case for all kinds of functors to our existing
solution, which up to now was entirely based on explicit specialisations.
C++11 supplied the new language construct 'decltype(EXPR)', which allows us
to capture any class with an function operator, which also includes the Lambdas.
The solution was proposed 2011 on StackOverflow
http://stackoverflow.com/questions/7943525/is-it-possible-to-figure-out-the-parameter-type-and-return-type-of-a-lambda/7943765#7943765
We used it already with success within our TreeMutator.
But obviously the goal should be to unite all the function trait / metaprogramming helpers,
which unfortunately is a more expensive undertaking, since it also involves
to get rid of the explicit specialisations and retrofit our Types<XXX...> helper
to rely on variadic templates rather than on loki-style typelists.
This first step here is rather conservative, since we'll still rely on our
explicit specialisations in most cases. Only the Lambdas will go through the
new, generic case, and from there invoke the specialisation for member functions.
The latter need to be rectified as well, which is subject of the next changeset...
Writing and debugging such tests is always an interesting challenge...
Fortunately this exercise didn't unveil any problem in the newly written
code, only some insidious problems in the test fixture itself. Which
again highlights the necessity, that each *command instance* needs
to be an independent clone from the original *command prototype*,
since argument binding messages and trigger messages can appear
in arbitrary order.
This is a little bit of functionality needed again and again;
first I thought to use the TypedCounter, but this would be overkill,
since we do not actually need different instances, and we do not need
to select by type when incrementing the counter. In fact, we do not
even need anything beyond just allocating a number.
So I made a new class, which can be used RAII style
this was a spin-off activity from writing the SessionCommand
function(integration) test, where I noted that we can't just
capture "a time value" as command memento
basically this is not necessary, since the compiler figures out
to use the conversion to target type when attempting to resolve
an equality comparison. But it helps to avoid ambiguities in cases
where several conversion paths do exist, e.g. when comparing string
with C-string
explicitly observed with the debugger that the call path is sane;
the code looks innocuous, but it is quite magic how the compiler
picks precisely the right ctors and inserts conversions apropriately
From a purely logical viewpoint, it looked sensible to require an actual
value for an offset, especially since our time values are immutable.
But this has the unfortunate consequence that we'd be unable to use
an offset value as parameter for any command, since we store the arguments
as tuple and the tuple type has a default constructor. We might be able
to get around that problem, but such looks brittle to me; it is just
plain surprising for anyone not familiar with the internals of the
command system.
For that reason, I've now added a default ctor to the Offset type
...since the session loop will be notified on any change via the
interface, adding a command will activate the loop, and the builder
timeout is handled separately via the dirty state. So there is no
need to spin around the loop in idle state.
As a aside, timeout waiting on a condition variable can be intentional
and should thus not be logged as an error automatically. It is up to the
calling context to decide if a timeout constitutes an exceptional situation.
It is always a trade-off performance vs. readability.
Sometimes a single-threaded implementation of self-contained logic
is preferable to a slightly more performant yet obscure implementation
based on our threadpool and scheduler.
Did a full review of state and locking logic, seems airtight now.
- command processing itself is unimplemented, we log a TODO message for now
- likewise, builder is not implemented
- need to add the deadlock safeguard #1054
And yes, this warning is for real, while the compiler has no way
to decide if there is actual danger lurking. A type with internal
linkage (e.g. defined in an anonymous namespace) will be treated
by the linker as a separate entity on each encounter (i.e. in
each distinct compilation unit). When multiple translation units
start collaborating on such a type, they *might* be referring
to different memory locations, while semantically the intention
is to refer to the same location.
And since we're dealing with a library facility here, *we* have
likewise now power to ensure proper usage, so we better be cautious.
after reading some related code, I am leaning towards a design
to mirror the way command messages are sent over the UI-Bus.
Unfortunately this pretty much abandons the possibility to
invoke these operations from a client written in C or any
other hand made language binding. Which pretty much confirms
my initial reservation towards such an excessively open
and generic interface system.
...the sheer amount of mechanical replacements scattered all over these
files might be a vivid indication, that the design of the interface system
is subobptimal ;-)
Phew, convoluted.
And I was doubtful that we need to support multiple typed child collection
Well, we get three such collections already in the first real world example...
reason is, only files with a @file comment will be processed
with further documentation commands. For this reason, our Doxygen
documentation is lacking a lot of entries.
HOWTO:
find src -type f \( -name '*.cpp' -or -name '*.hpp' \) -not -exec egrep -q '\*.+@file' {} \; -print -exec sed -i -r -e'\_\*/_,$ { 1,+0 a\
\
\
/** @file §§§\
** TODO §§§\
*/
}' {} \;
Damn sideeffect of the suppport for move-only types: since we're
moving our binding now into place /after/ construction, in some cases
the end() iterator (embedded in RangeIter) becomes invalid. Indeed this
was always broken, but didn't hurt, as long as we only used vectors.
Solution: use a dedicated init() hook, which needs to be invoked
*after* the TreeMutator has been constructed and moved into the final
location in the stack buffer.
unintentionally we used copy construction in the builder expression,
wenn passing in the CollectionBinding to the ChildCollectionMutator.
The problem is that CollectionBinding owns a shaddow buffer, where
the contents of the target collection are moved temporarily while
applying the diff. The standard implementation of copy construction
would cause a copy of that shaddow buffer, which boils down to
a copy of the storage of the target collection.
If we want to support move-only types in the collection, most notably
std::unique_ptr, we can thus only use the move constructor. Beyond that
there is no problem, since we're only ever moving elements, and new
elements will be move constructed via emplace() or emplace_back()
actually this is a pragmatic extension for some special use cases,
and in general rather discurraged, since it contradicts the
established diff semantics. Yet with some precaution, it should
be possible to transport information via an intermediary ETD
Map -> ETD -> Map
for the record: while it is indeed sweet-and-simple to support Ref::THIS
here, it is near impossible to represent it in general, in a setup with
multiple "onion-layers". The reason is, we'd have to incorporate such
special treatment into the /selector predicate/, which in turn undermines
the ability to pick the right onion layer to handle a given diff verb,
since "Ref::THIS" is a generic marker and we have no other data to base
the decision in the selector on.
Up to now, InPlaceBuffer used to default construct an instance of the
Interface class, and then you'd need to invoke the `create()` function
to actually create the desired subclass. This is not only inefficient,
but rules out the use of abstract interfaces / base classes.
Unfortunately, there is no way in C++ to specify an explicit template argument list
on ctor calls, so we resort to the trick of passing an additional dummy marker argument
yay! this piece of code has served its purpose:
it was the blueprint to build a way better design and implementation,
which can now cover this "generic tree" case as a special case as well
this adds kind of an extension point to diff::Record<GenNode>::Mutator,
which is then actually defined (implemented) within the diff framework.
This allows the TreeDiffTraits automatically to use this function
to get a TreeMutator for a given Rec::Mutator. Which in turn allows
the generic version of DiffApplicator automatically to attach and
bind to a Record<GenNode>
together this allows us to ditch the explicit specialisation
and dedicated, hand-written implementation of DiffApplication
to GenNode in favour of using the TreeMutator and friends.
this is a subtle change in the semantics of the diff language,
actually IMHO a change towards the better. It was prompted by the
desire to integrate diff application onto GenNode-trees into the
implementation framework based on TreeMutator, and do away with
the dedicated implementation.
Now it is a matter of the *selector* to decide if a given layer
is responsible for "attributes". If so, then *all* elements within
this layer count as "attribute" and an after(Ref::ATTRIBS) verb
will fast forward behind *the end of this layer*
Note that the meta token Ref::ATTRIBS is a named GenNode,
and thus trivially responds to isNamed() == true
needed to use a forward function declaration within the
lambda for recursive scope mutator building, since otherwise
everything is inline and thus the compilation fails when it
comes to deducing the auto return type of the builder.
Other than that, the whole mechanics seem to work out of the box!
previously they where included in the middle of tree-mutator.hpp
This was straight forward, since the builder relies on the classes
defined in the detail headers.
However, the GenNode-binding needs to use a specifically configured
collection binding, and this in turn requires writing a recursive
lambda to deal with nested scopes. This gets us into trouble with
circular definition dependencies.
As a workaround we now only *declare* the DSL builder functions
in the tree-mutator-builder object, and additionally use auto on
all return types. This allows us to spell out the complete builder
definition, without mentioning any of the implementation classes.
Obviously, the detail headers have then to be included *after*
the builder definition, at bottom of tree-mutator.hpp
This also allows us to turn these implementation headers into
completely normal headers, with namespaces and transitive #includes
In the end, the whole setup looks much more "innocent" now.
But beware: the #include of the implementation headers at bottom
of tree-mutator.hpp needs to be given in reverse dependency order,
due to the circular inclusion (back to tree-mutator.hpp) in
conjunction with the inclusion guards!
...instead of using a hand written implementation,
the idea is to rely on the now implemented building blocks,
with just some custom closures to make it work.
- esp. verify the proper inclusion of the Selector closure in all Operations
- straighten the implementation of Attribute binding
- clean-up the error checking helpers
similar reordering for the third part.
This time most operations are either passed down anyway,
or are NOP, since attribute binding has no notion of 'order'
yay! unit testing rocks.
Actually I changed the test definition for another reason, just to discover
that I've missed to implement that operation in this onion layer
now failing due to a contradiction in test fixture:
it is nonsensical to re-order attributes; rather, we should
cover re-ordering of children, to verify that the mutator binding
properly surpasses the attribute layers and forwards operations
to the lower layers responsible for handling child scopes...
In Theory, acceptSrc and skipSrc are to operate symmetrically,
with the sole difference that skipSrc does not move anything
into the new content.
BUT, since skipSrc is also used to implement the `skip` verb,
which serves to discard garbage left back by a preceeding `find`,
we cannot touch the data found in the src position without risk
of SEGFAULT. For this reason, there is a dedicated matchSrc operation,
which shall be used to generate the verification step to properly
implement the `del` verb.
I've spent quite some time to verify the logic of predicate evaluation.
It seems to be OK: whenever the SELECTOR applies, then we'll perform
the local match, and then also we'll perform the skipSrc. Otherwise,
we'll delegate both operations likewise to the next lower layer,
without touching anything here.
--> now it becomes obvious that we've mostly
missed to integrate the Selector predicate properly
in most bindings defined thus far. Which now causes
the sub-object binding to kick in, while actually
the sub-value collection should have handled
the nested values CHILD_B and CHILD_T
OMG, this is intricate stuff....
Questionable if anyone (other than myself) will be able
to get those bindings right???
Probably we'll need yet another abstraction layer to handle
the most common binding situations automatically, so that people
can use the diff framework without intricate knowledge of
TreeMutator construction.
- an extension to our custom toString and typeString helpers.
- currently just for shared_ptr and unique_ptr
- might add further overloads for other smart-ptr types
integrated into the generic DiffApplicationStrategy.
The dedicated, explicit specialisation for DiffMutable is
no longer needed, since the generic template will degrade or
fall back to precisely this functionality, when the target
implements the DiffMutable interface
This is the first skeleton to combine all the building blocks,
and it passes compilation, while of course most of the binding
implementation still needs to be filled in...
It occurred to me, that 90% of this template specialisation
are entirely generic and not dependant on the actual target type.
While the compiler/linker is able to sort such a situation out,
this might lead to template bloat and possibly subtle errors.
So it seems more adequate to emit the generic part of the code
right away from within a dedicated translation unit within the
library module; so the vtable is already in place and only
the flexible part of the code needs to be re-emitted on
each usage site.
- default recommendation is to implement DiffMutable interface
- ability to pick up similar non-virtual method on target
- for anything else client shall provide free function mutatorBinding(subject)
PERSONAL NOTE: this is the first commit after an extended leave,
where I was in hospital to get an abdominal cancer removed.
Right now it looks like surgery was successful.
this is at the core of the integration problem: how do we expose
the ability of some opaque data structure to create a TreeMutator?
The idea is
- to use a marker/capability interface
- to use template specialisation to fabricate an instance of that interface
based on the given access point to the opaque data structure
but unfortunately this runs straight into a tough problem,
which I tried to avoid and circumvent all the time:
At some point, we're bound to reveal the concrete type
of the Mutator -- at least to such an extent that we're
able to determine the size of an allocator buffer.
Moreover, by the design chosen thus far, the active
TreeMutator instance (subclass) is assumed to live within
the top-level of a Stack, which means that we need to
place-construct it into that location. Thus, either
we know the type, or we need to move it into place.
initially, even the diff applicator was meant to be a
"throwaway" object. But then, on writing some tests,
it seemed natural to allow re-using a single applicator,
after having attached it to some target.
With that change, I failed to care for the garbage
left back in the "old" sequence after applying one diff;
since in the typical usage sequence, the first use builds
content from scratch, this problem starts to show up only
with the third usage, where the garbage left from the input
of the second usage appears at the begin of the "new sequence"
Solution is to throw away that garbage explicitly on re-entrance
..because actually we don't know if the intention is
to drop those waste elements -- and for sure this
discarding of waste does not happen through the
invocation logged here; rather it happens by
abandoning the scope
...which mostly just is either ignoring the
operations or indicating failure on attempt to
'reorder' attributes (which don't have any notion of 'ordering')
overall, the structure of this implementation is still rather confusing,
yet any alternatives seem even less convincing
- if we want to avoid the delegation to base-class, we'd have
to duplicate several functions and the combined class would
handle two distinct concerns.
- any attempt to handle the IDs more "symmetrically" seems to
create additional problems on one side or the other
this also supersedes and removes the initial implementation
draft for attribute binding with the 'setAttribute' API
The elementary part of diff application incl. setting
new attribute values works by now.
While in general it is fine to clean-up any entity IDs
to be US-ASCII alphanumerics (plus some allowed interpunction),
the GenNodes and also keys in object-bindings for diff are
considerd internal interfaces, assuming that any passed
ID symbol is already sanitised and checked. So the
sanitise operation can be skipped. This changeset
adds the same option directly to lib::EntryID,
allowing to create an EntryID that matches
a similar GenNode's (hash) ID.
The way we build this attribute binding, there is no single
entity to handle all attribute bindings. Thus the only way
to detect a missing binding is when none of the binding layers
was able to handle a given INS verb
to summarise, it turned out that it is impossible to
provide an airtight 'emptySrc' implementation when binding
to object fields -- so we distinguish into positive and
negative tests, allowing to loosen the sanity check
only for the latter ones when binding to object fields.
..as concluded from the preceding analysis.
NOTE this entails a semantical change, since this
predicate is now only meant to be indicative, not conclusive
remarks: the actual implementation of the diff application process
as bound via the TreeMutator remains yet to be written...
how can ordinary object fields be treated as "Attributes"
and thus tied into the Diff framework defined thus far.
This turns out to be really tricky, even questionable
while simple to add into the implementation, this whole feature
seems rather qestionable to me now, thus I've added a Ticket
to be revisited later.
In a nutshell, right here, when implementing the binding layer
for STL collections, it is easy to enable the framework to treat
Ref::THIS properly, but the *actual implementation* will necessarily
be offloaded onto each and every concrete binding implementation.
Thus client code would have to add support for an rather obscure
shortcut within the Diff language. The only way to avoid this
would be to change the semantics of the "match"-lambda: if this
binding would rather be a back-translation of implementation data
into GenNode::ID values, then we'd be able to implement Ref::THIS
natively. But such an approach looks like a way inferiour deisgn
to me; having delegated the meaning of a "match" to the client
seems like an asset, since it is both natural and opens a lot
of flexibility, without adding complexity.
For that reason I tend to avoid that shortcut now, in the hope
to be able to drop it entirely from the language
...basically this worked right away and was easy to put together.
However, when considering how many components, indirections and
nested lambdas are working together here, I feel a bit dizzy...
:-/
write down a first draft for a definiton section,
to describe the fundamental parts involved, when
applying a diff message onto implementation defined
data structures
After a break of tree weeks, I found it difficult to find may way
amidst all those various levels of abstraction. In addition to this
definition, we'll probably also need a high level overview of the
whole diff system operation.
...all of this implementation boils down to slightly adjusting
the code written for the test-mutation-target. Insofar it pays off now
having implemented this diagnostic and demonstration first.
Moreover I'm implementing this basic scheme of "diff application"
roughly the fourth time, thus things kindof fall into place now.
What's really hard is all those layers of abstraction in between.
Lesson learned (after being off for three weeks, due to LAC and
other obligations): I really need to document the meaning of the
closures, and I need to document the "abstract operational semantics"
of diff application, otherwise no one will be able to provide
the correct closures.
