NilPy has no reachable path to the sys* and arg* intrinsics

• Type: feature (a capability gap) — Track N
• Found: 2026-09-05 (frankS), while deleting the dead arms that hid it
• Compiler: 3d613aac6251

What was measured

PyParseFactorCore carried five arms — tkSysOpen, tkSysRead, tkSyswrite, tkArgCount, tkArgStr — matching those spellings as tokens. Two things were true at once:

1. The arms could not fire. 5f177b181 made those spellings soft keywords, so they lex as plain tkIdent and the tk enum members survive only as -Ord(tkXxx) intrinsic call ids. Confirmed four ways that fail differently: frankD's dynamic probe, the 2026-09-04 reachability sweep over all 830 .npy programs, the lexer emitting no such token for either frontend, and the Pascal parsers already having zero such arms.
2. Every -Ord(tkSysOpen) … -Ord(tkArgStr) construction site in pyparser.inc was inside those arms. Five arms, five construction sites, all in the dead region.

So the arms were the only path to those intrinsics from NilPy, and the path was unreachable. NilPy has never been able to call them.

Why this is a ticket rather than a line in a commit

Deleting dead code is normally silent. Here it removes the only trace that NilPy was ever intended to have these, and a missing fact collides with nothing: without this ticket, the next person to ask "can NilPy call sysopen" finds no arms, no ids, no tests and no history, and concludes it was never wanted. That may even be right — but it should be a decision, not an absence.

The actual question, and it may well be "no"

Do not implement this reflexively. Two arguments against:

• A NilPy program can already declare and bind its own paramstr / paramcount — frankD measured precisely that, and it works today (verified again at 3d613aac6251: a .npy defining both prints 3 and 8).
• Python's idiom is sys.argv, not a ParamStr intrinsic. NilPy is upward compatible with CPython, so adding a Pascal-shaped intrinsic is the kind of thing nilpy-semantics-divergences.md exists to keep deliberate.

The argument for is raw file I/O without the pylib layer, which is a real want on constrained targets — but that is a use case someone should name, not an inference from five arms that never ran.