Modified TYPECHECK query to include "checkingScope" query parameter:
it is a root node for which program is generated (for local typechecking it's local node).
Rules additionally match on a constraint `checkDecls(node<>)` bearing this node,
so while rules will be generated for all scopes, only those in applicable scopes will run.
Also split rule for EnumConstantDeclaration on declration checking and just typing rule.
Now primitive types in conversion relation are handled before processing of free inference vars.
So primitive types first get boxed and only then these terms can be unified with inference vars.
To avoid confusion, a rule defines "target ref", which is supposed
to indicate the source code location for possible errors.
The origin node is to be referred as "origin" as before.
Temporarily rename OriginSpec's alias to "target" to reflect change
in semantics.
Introduce SessionData, allow to keep arbitrary data with the session.
Extend RulesListTable with support for invalidation based on rule unique tags.
Introduce UpdatesRecorder to keep track of invalidated nodes.
Extend ApplyTemplatesStep with invalidate logic for typechecking.
Simplify design, drop unnecessary fields, minor refactorings.
Introduced a variation of conversion relation with a bit stricter boxing, needed for eq/neq.
e.g. suppose { (Number n, float f) => n == f } here f shouldn't be boxed.
Everywhere where we don't care whether it's <String> or <string> it will be handled just as String.
Where we need to discern between the cases (as in Recover), 'stringType()' dataform can be matched first.
Remove rules which handle capture+inference var case separately: it's already handled.
Remove rule with error on bound which shouldn't be triggered with new hasBound semantics.
Add symmetric rule for case value+inference var: types can be inferred nevermind relation direction.
Add rules for free vars without bounds (these ain't inference vars). Need in such vars arise when
e.g. params in classifierType terms must be somehow accessed. It's used in MethodCall rules.
During investigation a different issue has been found.
Currently, e.g. when checking args to a method call, if there's an error then inference variable remains free.
It allows to continue typechecking without spurious additional errors triggered by former ones,
but with this approach all relations must gracefully handle free vars, and it's more difficult to avoid "no type".
There's an alternative to always assign to inference var a type at hand, even it doesn't pass checks
-- but then there're more unneccessary errors. Both approaches seem valid, no action is taken now.
Errors are now reported on ClassifierType instead of type argument, because some checks must be done
in captures and there's no proper way to specify origin of arg here.
Fedor Isakov