A bare method call inside its own class ignores its arity

Type: bug — Track P (compiler/pasparser_stmt.inc, the implicit-Self call site; helper in compiler/pasparser_lval.inc).
Found inside the candidate-selection group, while measuring whether [[bug-p-a-shadowed-soft-intrinsic-is-closed-without-consulting-the-arguments]] still had a live instance. It does not — but a class with a method named Delete shadowing the intrinsic turned out to accept Delete(a, 1, 1) and run Delete(Double) with 0.0. That has nothing to do with intrinsics.

The measurement

fpc 3.2.2 refuses every row; pre-fix pxx compiled all four clean.

call	signature	pxx ran
`Plain(1.5, 2, 3)`	`Plain(x: Double)`	`x=0.0` — every argument lost
`Plain(1.5, 2)`	`Plain(x: Double)`	`x=0.0`
`Two(7, 8, 9)`	`Two(a, b: Integer)`	`a=8 b=9` — Self ate the 7
`Two(7)`	`Two(a, b: Integer)`	`a=369098760` — uninitialised

The last row is why this is not a diagnostics ticket. There is no crash and no message; the callee reads whatever the argument slot held.

Exactly one door of four

spelling	pre-fix
`Self.Plain(1.5, 2, 3)`	refused
`c.Plain(1.5, 2, 3)`	refused
`Free1(1.5, 2, 3)` (free routine)	refused
`Plain(1.5, 2, 3)` bare, inside the class	accepted

FindUMethOverloadAhead is a selector — asked which overload, having already been told this is a method call — so with no arity-viable candidate it answers with the first name match rather than declining. FindUMethArityStrict exists precisely so a caller can decline, and its own header says so; it had one caller, the Write-inside-a-write-member path. This is the sibling site that was never given the strict question. Same shape as [[bug-p-a-method-call-with-missing-arguments-is-accepted-and-reads-garbage]], which fixed the QUALIFIED parenless spelling.

The carve-out is a measured dependency, not caution

Variadic bracket-elision (Desc('a', 1) against Desc(const a: array of const), feature-writeln-as-library) passes more explicit arguments than the signature has parameters, on purpose — and it reaches this site through the very loose fallback being closed. fpc refuses that source too ("Wrong number of parameters"), so an arity gate derived from fpc's answer alone would have deleted a pxx extension with no test to notice. Hence UMethNameCanAbsorbVarRecTail, asked of the NAME over the visible set rather than of one candidate — the same shape and the same reason as ParamIsVarRecArray beside it.

Not fixed here, and deliberately

The elided spelling of that extension segfaults at this door while the bracketed one works. Separate defect, separate ticket: [[bug-p-a-bare-variadic-method-call-segfaults-where-the-bracketed-spelling-works]]. Converting a segfault into a diagnostic as a side effect of an arity fix would have hidden it.

Verification

test_p_a_bare_method_call_ignores_arity_fail.pas — four diagnostics, rc=1, no binary.
test_p_a_bare_method_call_arity_still_valid.pas — exact arity, trailing defaults, a parameterless method, two overloads, an inherited method and the bracketed array of const. Byte-identical against the PINNED pre-fix compiler, so the check refuses nothing that was legal.
gate.sh quick GREEN, self-host fixedpoint converged.