The output can overwrite the source under another spelling

pxx g.pas ./g.pas      # exit 0, g.pas is now an ELF

Resolution (2026-09-28)

Compared by IDENTITY: compiler/output_guard.inc reads mnt_id and ino from /proc/self/fdinfo of an O_PATH descriptor (there is no stat intrinsic; the compiler already reads /proc through sysopen/sysread). The driver records the output's identity once, when the output already exists, and refuses when the source is it. Every other input read goes through LoadInput (LoadFile plus the check): Pascal units and includes, C sources and headers, objects and libraries read for linking, resources. LoadFile itself is an intrinsic in a self-hosted compiler, so the check cannot live in its body.

A NAMED output with the source's exact spelling is now refused as well (it was renamed to <source>.out); a DEFAULTED output equal to the source (pxx g, no extension) keeps the .out rename. Without /proc the spelling check remains, as before.

Not covered: elfdynsym.inc's .dynsym probe of a candidate shared library (test_elfdynsym.pas includes that file standalone, without the guard), and side artefacts beside the output (a map file) -- neither is a named output.

Fixed with it: --no-lazy-var's error told the user to pass --lazy-var, an unknown option; it now names the option that disabled the feature.

Fixture test/the_output_may_not_be_an_input.sh (13 rows; test-core): v445 fails 11, this build none. Cost measured on the self-host compile with an existing output: 204 extra opens, no measurable time.