A Char array through a FIELD or a DEREF is not a string

Found by attempting the target: three lib-test lib_synapse reds were wired to [[umbrella-one-full-tier-run-with-no-red-tier]] as unreproducible because external/synapse is absent on plexus. Fetched it; all three reproduce, and all three are one construct.

The construct

external/synapse/ssfpc.inc, inside WSAStartup's with WSData do:

szDescription  := 'Synsock - Synapse Platform Independent Socket Layer';
szSystemStatus := 'Running on Unix/Linux by FreePascal';

Both fields are array[0..N] of Char in TWSAData. Two lines, two errors, and the error is the same one in all three jobs:

pascal26:0: error: incompatible types: cannot assign ShortString to Char

Root cause — one oracle, one node kind

ASTCharArrayCap (pasparser_lval.inc) is what every consumer of "a static array of Char IS a string" asks: the AN_ASSIGN arm in ir.inc for the store, WrapCharArrayToStringExpr for the read, the argument coercion, the comparison. Its header said "the element count if node denotes a STATIC Char array". Its body said:

if ASTKind[node] <> AN_IDENT then Exit;

A field, an element or a deref answered -1, the conversion never fired, and the assignment fell through to AssignSideKind's AN_FIELD arm — which reads ASTTk, and ASTTk on an array field is the ELEMENT kind. So the destination typed as Char, the source as ShortString, and the check refused it. The check was right; it was answering the question it was given.

Measured before the fix, seven lvalue shapes:

Five of six, and the exact mirror of the defect AssignSideKind's own header records, where five of six lvalue shapes were UNCHECKED and a string handle was stored over a record. Same function, same six shapes, opposite sign. A predicate that answers about an IDENT while its name says NODE is wrong in whichever direction its caller happens to face — and both faces have now been paid for.

The fix

ASTCharArrayCap learns AN_FIELD and AN_DEREF, through the accessors that already know each shape (RecFieldIsArray/ArrLen/DynDepth/ArrNDims/ Type, and IsDerefPtrArray/DerefPtrArrayInfo) rather than by re-deriving the layout. IsNodeArray is the sibling predicate and asks exactly these three node kinds for exactly this reason — its own header names "writing to one was rejected as cannot assign ShortString to Char" as one of four faces of one wrong answer. This is a fifth face of the same one.

Multi-dimensional returns -1 in every arm rather than guessing an extent: a wrong capacity here is a write past the row, and a loud refusal is the better failure.

Verification

• test/test_char_array_field_is_a_string.pas, new, wired into test-core. 14 assertions covering all six fixed shapes in BOTH directions, plus the two things a compile-only test would miss: the zero fill (a shorter value over a longer one leaves ab......, not abcdefgh with two bytes patched) and truncation at cap, each with a neighbouring field asserted intact.
• The oracle is FPC 3.2.2, run on this same file: all 14 rows identical.
• Positive control: the pinned v399 compiler (954adef93a7b0e9e) REJECTS the test at three lines. A green run therefore cannot be a run that did nothing.
• All four synapse programs build and produce their Makefile-expected output byte for byte, including lib_synapse_tls_loopback's real handshake with the verify-rejects arm.
• converged after 1 round(s), binary 090042338fc2deae.

What this does NOT close

The three regression-lib-test-lib-synapse-* jobs build with $(PXX_STABLE), and the pinned compiler still has the bug. They stay RED until the owner takes the next pin, and no amount of work in this tree changes that. Stated on each of the three tickets.

Log

• 2026-09-02 — resolved, commit 9c6b216aa.