NilPy string subscripts are 1-BASED — silently off by one vs CPython

Found 2026-07-19 while landing [[feature-nilpy-str-methods]] unit 2. Pre-existing; not caused by that change. Silent wrong values, no diagnostic — which is why this is filed urgent rather than backlog.

Repro

s = "Hello"
print(s[0])
print(s[1])
print(s[4])

The subscript is being handed straight to Pascal's 1-based string indexing: index 0 reads the slot before the first character (a NUL here) and every other index is the previous character. s[len(s)-1] therefore reads the second-to-last char, and the last char is unreachable.

Why it matters for milestone 1

uforth uses 123 slice/subscript sites. Every one of them is silently wrong today, and wrong in a way that still produces plausible-looking output — a Forth tokenizer indexing word names one character early does not crash, it mis-parses. Any .npy corpus result involving string indexing is untrustworthy until this is fixed.

Note TPyList already implements Python semantics INCLUDING negative indices (see pylib's PyListFix); only the string path was missed. That asymmetry is the tell — the fix should route string subscripts through the same convention, negative indices included (s[-1] is the last character in Python).

Scope

Also settle slices (s[a:b]) while here — uforth leans on them heavily, and they share the convention. Check whether for c in s iterates correctly too.

Fixed 2026-07-19

s[i] in PyExprMode now lowers to a pylib pystr_at(s, i) call instead of a raw Pascal AN_INDEX. Doing the fixup in the RUNTIME helper rather than as index arithmetic in the parser keeps the negative case correct without evaluating the base twice (s[-1] needs Length(s)).

Semantics now match CPython: 0-based, negative counts from the end, and out of range prints IndexError: string index out of range and halts — the same shape as TPyList's existing behaviour.

Two files encoded the OLD behaviour and were ported

• test/test_nilpy_str_param.npy asserted "ab"[1] == "a". Python says "b". Its scanning loop was 1-based (i = 1; while i <= n + 1); now i = 0; while i <= n. Expectation updated 2\na\ncd\nok! -> 2\nb\ncd\nok!.
• examples/shell/shell0.npy (Track B/E file, ported out of necessity — the semantics change would otherwise leave the demo broken): four index sites in tok, ap_wc, ap_upper, ap_rev. ap_rev went from i = Length(args); while i >= 1 to i = Length(args) - 1; while i >= 0.

That these two were the ONLY breakage is a useful signal about how little .npy code depended on the old convention.

Gate: test-nilpy GREEN, --tier quick GREEN, self-host byte-identical, test_nilpy_str_methods.npy (now including subscripts) diffed against CPython.

Follow-up filed: [[bug-nilpy-subscript-on-literal]].