Repro

$ printf 'import "/usr/include/GL/gl.h"\nprint(glGetError())\n' > g.py
$ ./compiler/pascal26 g.py out          # compiles; two host-header warnings only
$ ./out
./out: error while loading shared libraries: libgl.so: cannot open shared object file

On disk: /usr/lib/x86_64-linux-gnu/libGL.so, libGL.so.1, libGL.so.1.7.0.

Note import "/usr/include/GL/gl.h" with only print("reached") runs fine — the missing library is only reached when a symbol from it is actually called, so the trivial probe passes and the real one fails.

Why this is more than a tolower bug

sqlite3.h -> libsqlite3.so.0 works, which is what makes the rule look sound: the stem happens to be lowercase there, so the sample that confirms the rule is the one where case cannot matter. A header stem is not a soname in general (GL/gl.h -> libGL.so.1, SDL2/SDL.h -> libSDL2-2.0.so.0), so the mapping wants the real mechanism — pkg-config, an explicit table, or a declared name — rather than a case fix.

Whatever replaces it, keep the failure at LINK time rather than exec time. A green compile that dies on startup is the worst of the three outcomes.

RESOLVED 2026-09-10 — and the ticket's own cause was wrong

What it actually was

LowerCase is in CSonameForStem, and it is not what loses the case. The stem handed to it is already lowercase, because it is a FILE NAME:

/usr/include/GL/gl.h   ->  GetFileBaseName  ->  "gl"
                            ^^                   ^^
                            the capitals are here, and they are dropped here

So the ticket's "Why this is more than a tolower bug" section was right for a better reason than it gave: a header stem is not a soname and there was never a case to preserve. Removing the LowerCase would have changed nothing.

The mechanism it names — pkg-config, a table, a declared name — turned out to be already present and simply asked the wrong question. CSonameForStem consults /etc/ld.so.cache FIRST and falls back to nine hardcoded entries; the cache is the real, host-derived, non-tabulated mapping the ticket wanted. It just could not answer libgl.so when what it holds is libGL.so.1.

The case-insensitive pass, and why it is safe

LdCacheFindSoname now scans the whole table and returns a case-insensitive hit only when no exact one was seen. An exact hit always wins, which is the property that makes this an addition rather than a policy change: nothing that resolves today can move.

Measured on this box, 2026-09-10:

So the fallback is unambiguous here and serves libGL, libGLU, libEGL, libFLAC, libICE, libLLVM, libOpenCL and 169 others — not one library. Where a host does hold two keys differing only in case and neither matches exactly, this takes the first in cache order; that is a guess, and it is strictly better than the current answer, which is a name that is in the cache at all.

The half that outlives this ticket

"Keep the failure at LINK time rather than exec time" was the ticket's other instruction, and it is the one that generalises: no case fix can map every header to every library, so the residual has to be diagnosed rather than guessed.

The check went into RegisterExternal, which already carries its sibling — the __pxx_* guard, for a SYMBOL no library can export — and describes itself as "the choke point where an unresolved external becomes a real import." The LIBRARY half was simply absent, and the reason is visible in the history: every instance was fixed on the PRODUCER side, one at a time.

• bug-c-a-header-reached-by-uses-discards-function-bodies-and-imports-them-instead removed one way to synthesise "a lib<header>.so that cannot exist, dead at load".
• bug-a-a-c-include-path-captures-a-pascal-uses-and-emits-a-dynamic-import removed another: "the binary carries a DT_NEEDED on libstrings.so and dies at load. Nothing in the build says so."

Two authors wrote that sentence about their own defect and neither added the check that would say so. That is the ABSENCE shape, not a drift: no path was wrong, a position was empty, and an empty position collides with nothing.

PROVENANCE decides membership, and the measurement is why

The obvious guard — check every soname against the cache — is measurably wrong. Censused 2026-09-10 across every external '...' clause in the tree:

test_c_lazycasing.pas says it in its own comment: "soname only — the loader finds it via LD_LIBRARY_PATH (hermetic, like test_c_argspill)". An absent-from- cache library is a deliberate, in-tree-tested configuration, and a whole- soname check would refuse all four.

What separates them from libgl.so is not the name, it is who wrote it. A name in an external clause carries intent the compiler must not second-guess; a name the compiler derived from a file name is its own guess, and checking your own guess against the host is not second-guessing anybody. So only synthesised names are recorded (CSynthMissingLibs), and the guard fires on those alone.

And it is recorded at import but consumed at RegisterExternal, because an import whose symbols are never referenced emits no DT_NEEDED and is a working program — the ticket says so itself ("import ...SDL.h with only print("reached") runs fine"). Diagnosing at import time would have been a guard firing on a program that does not have the defect. That is the negative control, and it is a test row.

Tests

The first two are hermetic — libnolib.so exists on no machine — so they assert the guard without depending on what this box has installed. The GL row asserts a RELATION rather than a constant: whatever soname comes out must be one this loader resolves. That carries to a box whose GL is libGL.so.2, and it still fails on the pre-fix answer, because libgl.so resolves nowhere — so the row is its own positive control and cannot pass on a revert.

Not done, and it is measured rather than deferred by feel

See feature-n-derive-a-header-s-library-from-its-directory-and-verify-it-against-the-library-s-own-dynsym. The one-line version: the DIRECTORY is a far better library key than the file stem (617 headers versus 76 on this box), and net/if.h -> libnet.so.9 is why it cannot be added as written — that answer is a real library, so it passes this guard and dies later on undefined symbol, which is worse than the loud failure this ticket just fixed.

Gate: gate.sh quick RED on one row, mimic_queue :: unknown type: TPyDeque — a pylib builtin no pin carries, unrelated to this change and unchanged by it. self-host fixedpoint, fpc seed compiles (forward decls) and testmgr --tier quick all PASS.

Log

• 2026-09-10 — resolved; this names the commit that carried the resolve, which is not always the one that carried the change — commit e53eff428.