[ddehilster]

Two changes shipping in one build as 3.7.13.

Fixes #698

---

1. Function names containing digits or underscores (#698)

A user function whose name contained a digit (Round2, Clamp01, Fixed2) or an underscore (Round_2) never compiled. The rule-file tokenizer built _LIT from a single _xALPHA token, so Round2 reached the function grammar as _LIT("Round") + _NUM(2) and never matched _FNCALLLIST <- _LIT \( in gram2.nlp. The diagnostic was [Warning: Missing ';'] + [Syntax error.] on a function header that is itself valid — which sends you looking at the previous statement instead of at the name.

Takes the fix the issue prefers — accept the digits — rather than just reporting a better error, since digits have always been legal in L() variable names.

bigtok.nlp gloms adjacent alpha/numeric/underscore tokens into one _LIT, but only when immediately followed by an unescaped ( — a function definition header or a call. That gate is what keeps the change contained: @RULES literals must write \( and key=(val) pairs put = before the paren, so element literals (covid19 stays two elements) and pair values tokenize exactly as before. The second element is [plus], so single-token names never reach the rule and the keyword rules (if/while/return/L/N/X/G/P/s) still win.

postrfa.cpp gives rfaname() a two-arg form that gathers covered text across a range of collects; the one-arg form is untouched. The name still starts at an _xALPHA, so identifiers Ifunc::genDecls emits into generated C++ stay legal.

Verified: all forms from the issue table now run (Round2, R2ound, x2Round, Round_2, Clamp01, Round_Two). parse-en-us final.tree is byte-identical to the CI baseline, and an A/B run of all 15 tutorial/nlpfix/corporate analyzers against the unpatched grammar shows zero drift.

Adds .github/workflows/tests/fnname-digits plus a step in the Windows and Linux builds next to the existing #669 regression. The step asserts out.txt rather than the exit code — the engine exits 0 even when the analyzer fails to build. Confirmed the step fails on the unpatched grammar.

---

2. NLP++ math functions

NLP++ had only logten, abs, mod, factorial and randomint. This adds the rounding family and the remaining core C/C++ math functions.

Function	Returns	Notes
floor(num)	NUM	round down, toward −∞
ceiling(num)	NUM	round up, toward +∞
round(num)	NUM	nearest whole number, halfway away from zero
truncate(num)	NUM	discard the fraction, toward zero
sqrt(num)	FLOAT	errors on a negative argument
pow(base,exp)	FLOAT	errors when the result is NaN or out of float range
log(num)	FLOAT	natural (base e) log, argument must be positive

The four rounding functions take a float and return a whole number, which is what callers actually want. floor and truncate deliberately differ on negatives.

Registered in the parallel funcs.h / func_defs.h tables (verified in sync — 340 entries each), declared in fn.h, dispatched and implemented in fn.cpp.

Compiled analyzers: codegen in iaction.cpp emits Arun::<fnname> for every builtin, so each function also gets Arun overloads in fnrun.cpp and declarations in Arun.h — including the four RFASem/float combinations for the two-argument pow. The global C functions are called with the :: qualifier so the new Arun members don't recurse into themselves.

Help: seven new pages under visualtext/Help/helps modeled on logten.htm, TOC entries in Help.hhc, and Table_of_Math_Functions.htm rebuilt. That table previously listed logten alone, so abs, factorial, mod and randomint are now documented there too.

Testing — built clean and verified against a live analyzer:

floor(3.7)     = 3     floor(-3.7)     = -4
ceiling(3.2)   = 4     ceiling(-3.7)   = -3
round(3.5)     = 4     round(-3.5)     = -4
round(3.4)     = 3
truncate(3.7)  = 3     truncate(-3.7)  = -3
sqrt(16)       = 4     sqrt(2)         = 1.41421
pow(2,10)      = 1024  pow(9,0.5)      = 3
log(2.718282)  = 1     logten(1000)    = 3

sqrt of a negative and log(0) are rejected rather than propagating NaN.

---

Version

NLP_ENGINE_VERSION in nlp/main.cpp bumped to 3.7.13 — that #define is what --version prints, since the generated Version() is commented out at the callsite. Confirmed nlp --version reports 3.7.13.

Notes for review

• Negative literals as arguments (e.g. floor(-3.2)) fail to parse. Pre-existing, not introduced here — abs(-5), mod(-7,3) and logten(-1.0) fail identically. Passing a negative variable works. Worth noting the existing abs help page documents abs(-7), which does not actually run.
• abs, mod and randomint have no Arun:: overloads, so they are broken in compiled analyzers today. Not fixed here to keep scope contained — happy to follow up.
• min/max deliberately omitted. NLP++ distinguishes NUM from FLOAT and these are the only functions in the set with a genuinely ambiguous return type (a float-based min(2,3) could print 2.000000). Worth deciding explicitly rather than picking silently.

Companion PRs: VisualText/vscode-nlp#1112 (editor support), VisualText/visualtext-files#197 (help docs)

🤖 Generated with Claude Code