The three faces, measured

Compiler before: 48c9f5942757 at 72b9578d5. After: 8b10e02e2029. Independently reproduced under pin v407 (095ef4811a5b), so it is not a local-tree artefact. fpc 3.2.2 is the oracle on every row.

The loud row prints

pascal26:9: error: no overload of F$21 matches these arguments
  argument types: (Integer)
  candidates:
    F$21(Integer, record)
  in: ./compiler/builtin/builtinheap.pas
  note: that unit is appended to every program by the compiler -- you did not write it.

(Integer) is the spliced capture with the routine's own parameters gone, because a bare read has none by construction. Both halves of that message point away from the source: F$21 is a name the program does not contain, and the in: line accuses a unit the programmer did not write and the note then says so, which reads as "the compiler's own header is broken" rather than "line 9 of your file".

Mechanism

ParseNestedRoutine (compiler/pasparser_decl.inc) lifts a nested routine to top level and rewrites body references to its own name, mangling F to F$<n>. A capturing routine gains hidden leading parameters, so a self-reference that is a CALL must gain the matching actuals — that arm was added by bug-nested-proc-sibling-call-unresolved symptom 2 and is correct.

The discriminator it used was "followed by :=": a result write if yes, a recursive call if no. There is a third reading, and it is the one the flip in pasparser_expr.inc had already established everywhere else — in objfpc/default a bare own-name VALUE READ is the result var. It is not followed by :=, so it took the call arm.

The two passes then could not correct each other. pasparser_expr.inc's own-name-read branch is guarded Tokens[TokPos].Kind <> tkLParen, and by the time it runs the hoist pass has already inserted (Top) — so the branch that knows the right answer excludes itself on the evidence the wrong pass planted.

Why the silent face is the bigger one

The arity error only exists because the routine has parameters of its own. A paramless capturing nested function gets F(Top) spliced onto a signature that is exactly F(Top) — a well-formed, correct-arity, infinitely recursive call. test_nested_fn_bare_own_name_read.pas therefore asserts VALUES on every row; a compiles/does-not-compile row passes on the version that segfaults.

[[a-guard-that-cannot-fail-is-not-a-guard]] in its assertion-class form: the population here is "nested functions that capture", and within it the refusing member is the one with parameters. A corpus finds that one and says nothing about the other, which is the same silent/refused asymmetry [[feature-pascal-corpus-expansion]] recorded for the set-literal wall.

What this narrows, deliberately

Inside a CAPTURING nested function, statement-position F; no longer self-recurses — write F(). That is not a new rule, it is the existing one reaching here: the paramless flip (bug-bare-function-name-call-vs-resultvar) already made F() the spelling for self-recursion everywhere else and a nested routine was silently exempt. Measured: fpc refuses that spelling outright (Illegal expression), so no fpc-compatible source loses anything. Nested PROCEDURES and {$mode delphi} keep the old behaviour and both have a control row.

The narrowing's cost, measured by someone else

My evidence that nothing used the removed spelling was the self-host fixedpoint, and that is a claim about compiler.pas and lib/rtl (a build input) — not about the tree. It reads as covering more than it does, which is the failure mode CLAUDE.md already names for Track P coverage: partial, and worse than none because it looks total. lib/pcl, examples/, test/ and tools/ were the residual and I did not measure them.

frankA did, 2026-09-06, unprompted: 2508 .pas/.inc/.pp swept for a nested function F whose own name appears as a bare F; — first at line start, then anywhere on a line (after then/else/do, or beside another statement, which is where the spelling actually hides). 33 raw hits, zero real ones:

• test_nested_routine_depth2_capture.pas — eleven C; sites, every one a nested PROCEDURE. Its single nested FUNCTION already writes C := C(n - 1) + n: explicit parens to recurse, C := for the result. Both arms this fix keeps.
• test_nested_routine_local_shadows_own_name.pas — five Inner; statement sites, all nested procedures; its one function Inner is read as seen := Inner, which is the defect fixed here rather than the spelling removed.
• The rest: x := F; value reads (the fix again), property/type/var noise, and lib/rtl/palthreadobj.pas:463's WaitFor; — TThread.Destroy calling its own method, matched only because the first pass could not tell an indented method declaration from a nested routine.

What that census still cannot see, stated because an unlabelled claim beside a measured one inherits its credibility: Pascal outside this repo, a name reached through with or a unit qualification, and anything under compiler/ — which is exactly where the fixedpoint IS the right instrument. Real outside source using the spelling is a compat item, not a reopening of this call.

How it was found, and the correction that matters

Reducing the pparser.pp:2670 PeekOper wall. The reduction n1.pas was recorded as reproducing that wall; it was reproducing THIS one. Both print no overload of <name>$<n> matches these arguments with a mangled name and a capture-shaped candidate list, and the wall's own construct (Result:=PeekOper, a SIBLING call) is one token away from the shape that reproduces. The reduction had drifted onto a different defect and the error message could not tell me — n1 compiles and matches fpc now, and pparser.pp:2670 is unchanged.

The discriminator that separated them, once I stopped varying the enclosing scope: writing the same body with Result := instead of the own name. That row is kept in the fixture for exactly that reason.

Log

• compiler/pasparser_decl.inc — the third arm, with the mode/kind carve-out.
• test/test_nested_fn_bare_own_name_read.pas + .expected — 8 rows, faces and controls; .expected is fpc 3.2.2's own output.
• test/test_nested_fn_bare_own_name_delphi.pas + .expected — the mode arm.
• Makefile — two test-core rows (test_nestown26, test_nestowndel26).
• All of the above landed together, commit fa7f75bec. gate.sh quick GREEN at that tree; make compiler/pascal26 converged after 1 round(s).