A unit's finalization section is silently never executed

• Type: bug — Track A (unit finalization emission; sibling of the fixed bug-unit-init-begin-form-not-executed).
• Found: 2026-07-22, while fixing the begin-form init bug (its ticket asked for finalization to be "checked and either working or ticketed" — it is not working).

Repro

unit ufin;
interface
procedure Touch;
implementation
procedure Touch; begin end;
initialization
  writeln('init runs');
finalization
  writeln('finalization runs');
end.

Program using it prints init runs / main done — the finalization line never appears. FPC prints finalization runs after the main body, in reverse dependency order.

Where

compiler/parser.inc implementation-scan loop: ParseInitializationSection stops at the finalization identifier and the scan's fall-through skips its tokens one by one. There is no exit-chain hook: fixing this needs (1) parsing the section into a synthesized __fini_<unit> proc like __init_<unit>, and (2) an exit path that runs registered finalizers in REVERSE init order — including on Halt, which is where FPC runs them too.

Why prio 40, not 60

Same silent-drop shape as the init bug, but finalization sections are far rarer in real code (none in-tree today), and the missed work is cleanup at process exit, which the OS mostly reclaims anyway. Escalate if a real consumer appears.

Acceptance

• The repro prints the finalization line after the main body.
• Reverse-dependency-order test (mirror of test_unit_init_begin_form).
• Runs on Halt(n) too.
• Gate: make test + self-host byte-identical.

Log

• 2026-07-22 — resolved, commit fa8b67ae.