[ddehilster]

Since engine 4.2.15 a negative strpiece position counts from the end. Code that computes a position as strlength(s) - k now goes wrong when s is shorter than k: the position comes out negative and wraps round to the end, where it used to run off the start. strpiece(",", 0, strlength(",") - 2) gives "," instead of "", and a trim loop written that way hung a production analyzer on a one-character value.

What changed

spec/KBFuncs.nlp

• StripEndDigits started its scan one character early. A one-letter name only came back right because -1 wrapped, and a name with no trailing digits lost its last letter. Rewritten so positions only count up from 0.

• SeparateConCount computed the base as strlength(name) - strlength(num) - 1. Its else branch was the no-op L("stop");, so it never stopped at a letter. It now takes only the trailing digits, with no subtraction that can go negative.

• JsonKBRecurse wrote each separator after a child:

  • An array followed by a plain sibling was not JSON (a stray }, and no comma after the ]).
  • Two arrays side by side merged into one.
  • An element with no attributes or children ended in a trailing comma.

  Separators now go out before each child. Every shape that was already valid JSON is written byte for byte as before. The debug write of hier.txt on every call is gone. New helper: JsonHasAttributes.

• The jsonwrite builtin copied these bugs on purpose. It's fixed the same way in jsonwrite writes valid, separate arrays (4.2.37) nlp-engine#796 (4.2.37), so JsonKB and jsonwrite still give identical output.

spec/UtilFuncs.nlp

• prosifyrec trimmed with strpiece(txt, 0, del - 1) in two branches without the del > 0 guard its third branch got in 2011. At del 0 the end is -1, which took the whole text instead of none of it.
• Every other computed position in spec/, misc/ and languages/ subtracts only 1 or sits behind a length test, so it can't wrap. Checked by hand: the morph helpers (behind len > 4), the suffix confidence code (behind len > 2, and unreachable), numth, the percentage formatter, and the word remover with its # Sanity. checks.

Help

• strpiece.md: what changed for computed positions, with the "," example and the -k idiom. Also that out-of-range positions return "" but write strpiece: Error. to err.log, except a start past the end, which doesn't log.
• jsonwrite.md and New_in_Current_Version.md (new 4.2.15 and 4.2.37 entries) describe both fixes.

Before and after

Run with a one-pass test analyzer calling the functions; old = current main, new = this branch.

Name	StripEndDigits old → new	SeparateConCount old → new
""	0 → 0	0 → 0
a	a → a (old was right only because -1 wrapped)	0 → 0
12	0 → 0	0 → 0
abc12	abc → abc	abc, 12 → abc, 12
item3	item → item	item, 3 → item, 3
item	ite → item	0 → 0
a1b2	a1b → a1b	a1, 12 → a1b, 2

The JSON shapes below were written by JsonKBRecurse and by jsonwrite. Old = engine 4.2.36 with current KBFuncs; new = the #796 engine with this KBFuncs.

Children	Old	New	Builtin == NLP++
item1 item2 name	invalid	valid	yes / yes
item1 item2 thing1 thing2	valid, one merged array	valid, two arrays	yes / yes
a1b1 a1b2	valid, array a1	valid, array a1b	yes / yes
name item1 item2	valid	valid, byte-identical to old	yes / yes
row1 (bare), row2→child, tail	invalid	valid	yes / yes
nested arrays, then a plain sibling	invalid	valid	yes / yes

No output change in real analyzers. I patched the KBFuncs copies in VisualText/analyzer-templates, analyzers, nlp-tutorials, nlpfix-analyzers and parse-en-us (companion PRs). Then I ran each analyzer that carries a copy on its inputs before and after: 18 analyzers, plus parse-en-us. All output was unchanged; only the analyzer's own path in init.log differed. With UtilFuncs, prosify gives the same result for every limit from 1 to 30. Help/check-help-docs.ps1 passes; its one orphan, New_in_Current_Version, was already missing from index.md.

🤖 Generated with Claude Code