Compiled analyzers can call interpreter-only builtins (4.2.25)

[ddehilster]

Problem. The code generator writes every builtin call as Arun::<name>(...). A builtin added to the interpreter (Fn::) without a compiled Arun:: twin therefore made every compiled analyzer that calls it fail to build (C2039: '<name>': is not a member of 'Arun'). An audit compiled one pass per builtin and found 23 in that state:

Arrays	arrayreverse arrayslice arraysort arrayunique push
Strings	strindexof strlastindexof striscaps strjoin
Regex	rematch refind resubst
Files	readfile readlines fileexists direxists dirlist filesize deletefile
Other	abs randomint jsonparse jsonwrite

jsonwrite matters most. The KB template's KBFuncs.nlp calls it, so KB-template analyzers such as date-time could not be compiled at all.

Fix. For these builtins the generator now emits Arun::fnbridge(nlppp, "name", {args}):

• The list is a bridgedFns table in Iaction::genEval.
• fnbridge (new lite/arunbridge.cpp) wraps the evaluated arguments as a temporary action and calls Fn::fnCall, exactly as Iaction::eval does, then frees them the same way.
• Compiled and interpreted analyzers therefore run the same code, and nothing is reimplemented.
• Arguments go in a braced list, which C++ evaluates left to right. That keeps the ordering guarantee the generator already maintains for multi-argument calls.

Builtins that do have an Arun:: twin are unchanged.

Testing

• New fixture tests/compiled-bridge, one pass per bridged builtin, each printing its result. It now compiles, and the compiled output is identical to the interpreted output (23 lines).
• I compiled the whole date-time analyzer, which uses jsonwrite through the KB template. It builds and runs, and its output.json, final.tree and datetimes.kbb are identical to the interpreted run, with no errors.
• New Windows CI step after "Package compile libraries". It runs -COMPILEANA on the fixture, builds it with a checked-in CMake template against the libraries just packaged, runs it with -COMPILED, and diffs the output against the interpreted run. A builtin added later without compiled support now fails CI instead of breaking users' compiled analyzers.
• All other Windows CI steps (39) pass locally, the -DEV Full English Test output is unchanged, and all 20 analyzers (21 inputs) produce byte-identical output to 4.2.18. The new step only runs in CI, because it needs the packaged compile-libs. I ran its commands by hand locally against my own engine libraries, with identical output.

Note. adoptconcepts and makeparentconcept are in the builtin table, but their implementations are inside #ifdef WORDNET. So in normal builds they're "unknown function" even in the interpreter. They're undocumented; this PR leaves them alone.

Merge after #779. Both bump nlp/main.cpp; this one is 4.2.25.

🤖 Generated with Claude Code