recipes ask whether a header is there instead of inferring it from the platform

Two recipes decide their configuration by branching on the PLATFORM and then
assuming what that platform's userspace carries. Both assumptions held on
every target this index had until a C library began presenting a different
environment on one of them, and the compatibility measurement has been
recording the result as a member failure.

CURL, ON LINUX. `#define HAVE_LINUX_TCP_H 1` sits inside
`#if defined(__linux__)`. openkal runs on the Linux kernel, so the predicate
is CORRECT; what is wrong is reading it as "glibc's userspace headers are
installed". Measured chain, each step taken after the previous one was fixed:

  1. `lib/setopt.c:31: fatal error: 'linux/tcp.h' file not found`
     -> ask `__has_include(<linux/tcp.h>)`. Same for `sys/eventfd.h`,
        `termio.h`, `sys/xattr.h`.
  2. `strerr.c:30: error: "strerror_r MUST be either POSIX, glibc style"`
     -> the glibc flavour was correctly withdrawn for musl and nothing
        replaced it; curl refuses a configuration naming neither. The POSIX
        flavour is declared when `__GLIBC__` is absent.
  3. `the C library in this graph (musl) declares that it does not supply the
     following, and the link has just asked for it: eventfd as timerfd_create`
     -> A HEADER TEST IS NOT A SYMBOL TEST. openkal-musl carries
        `sys/eventfd.h` and declares `eventfd` absent, so `HAVE_EVENTFD` must
        not follow from the header being present. Steps 2 and 3 were defects
        this change introduced and the next build caught; they are recorded
        because the distinction they turn on is the point.
  4. `undefined symbol: __isoc23_strtol`, referenced from
     `compat-x-openssl/3.5.1/lib/libcrypto.a`.

STEP 4 IS NOT A RECIPE DEFECT AND IS NAMED IN `compat.openssl.lua`'s OWN
COMMENT: "Because that build runs OUTSIDE mcpp's compile rules, it inherits
none of the resolved toolchain's flags --- it just calls `cc`." That archive
is built against the host's glibc and cannot link over openkal. curl on linux
is blocked on a packaging property, not on the C library or the engine.

C-ARES, ON WINDOWS. Its Windows branch defines `HAVE_WINDOWS_H`,
`HAVE_WINSOCK2_H`, `HAVE_STRCMPI`, `HAVE_STRNICMP` and no sub-second clock,
because until openkal every PE target had Win32. Each was measured:

  1. `ares_setup.h:81: fatal error: 'windows.h' file not found`
     -> `__has_include` for each of the four Windows headers.
  2. `call to undeclared function 'strcmpi'` / `'strnicmp'`, and
     `ares_timeval.c:93: error: missing sub-second time retrieval function`
     -> the CRT spellings are the ones absent here; `strings.h` decides, and
        `clock_gettime` is declared where `_WIN32` is not.
  3. `incomplete definition of type 'struct hostent'`.

AND STEP 3 IS WHERE ENTRY-BY-ENTRY PATCHING STOPS BEING THE RIGHT SHAPE. Each
fix moved the failure to the next Win32 assumption in the same block. That
block is a Win32 CONFIGURATION; a PE target presenting POSIX needs the POSIX
one, not a patched Win32 one. Choosing between them is a question about the
C environment --- `cfg(c-abi = ...)` in mcpp's vocabulary --- and the
`generated_files` selection is keyed by OS. The four fixes here are correct in
themselves and are kept for that reason; they are not sufficient, and this
records why rather than leaving the next reader to re-derive it from the
diagnostics.

Neither member reaches `runs` with this change. What it buys is that every
failure now names something true.

Author: Sunrisepeak

pkgs/c/compat.c-ares.lua

@@ -1030,7 +1030,9 @@ package = {
 #define HAVE_BOOL_T
 
 /* Define to 1 if you have the clock_gettime function and monotonic timer. */
-/* #undef HAVE_CLOCK_GETTIME_MONOTONIC */
+#if __has_include(<time.h>) && !defined(_WIN32)
+#define HAVE_CLOCK_GETTIME_MONOTONIC 1
+#endif
 
 /* Define to 1 if you have the closesocket function. */
 #define HAVE_CLOSESOCKET
@@ -1087,7 +1089,12 @@ package = {
 /* #undef HAVE_GETSERVBYPORT_R */
 
 /* Define to 1 if you have the `gettimeofday' function. */
-/* #undef HAVE_GETTIMEOFDAY */
+/* SUB-SECOND TIME. `ares_timeval.c:93` refuses a configuration that names no
+ * source of it, and the Win32 one (`GetTickCount64`) needs `windows.h`. A
+ * POSIX-presenting target has `clock_gettime`, which `time.h` declares. */
+#if __has_include(<time.h>) && !defined(_WIN32)
+#define HAVE_GETTIMEOFDAY 1
+#endif
 
 /* Define to 1 if you have the `if_indextoname' function. */
 /* #undef HAVE_IF_INDEXTONAME */
@@ -1205,10 +1212,22 @@ package = {
 #define HAVE_STDLIB_H
 
 /* Define to 1 if you have the strcasecmp function. */
-/* #undef HAVE_STRCASECMP */
+/* WIN32 CRT NAMES vs POSIX NAMES, decided by what the C environment
+ * presents rather than by the platform. This block is the Windows branch of a
+ * recipe that branches on the PLATFORM; over openkal the target is PE and the
+ * C environment is POSIX, so the CRT spellings are the ones that are absent.
+ * `__has_builtin` cannot answer for library functions, so these ask the same
+ * question the headers do: `strings.h` carries the POSIX pair. */
+#if __has_include(<strings.h>)
+#define HAVE_STRCASECMP 1
+#define HAVE_STRNCASECMP 1
+#endif
+/* #undef HAVE_STRCASECMP_ORIGINAL */
 
 /* Define to 1 if you have the strcmpi function. */
+#if !__has_include(<strings.h>)
 #define HAVE_STRCMPI
+#endif
 
 /* Define to 1 if you have the strdup function. */
 #define HAVE_STRDUP
@@ -1229,7 +1248,9 @@ package = {
 /* #undef HAVE_STRNCMPI */
 
 /* Define to 1 if you have the strnicmp function. */
+#if !__has_include(<strings.h>)
 #define HAVE_STRNICMP
+#endif
 
 /* Define to 1 if you have the <stropts.h> header file. */
 /* #undef HAVE_STROPTS_H */
@@ -1276,23 +1297,42 @@ package = {
 /* Define to 1 if you have the <time.h> header file. */
 #define HAVE_TIME_H
 
-/* Define to 1 if you have the <unistd.h> header file. */
-/* #undef HAVE_UNISTD_H */
-
-/* Define to 1 if you have the windows.h header file. */
+/* THESE ASK WHETHER A HEADER IS THERE, SO THEY ASK THE COMPILER.
+ *
+ * This block is the Windows branch of a recipe that branches on the PLATFORM,
+ * and it used to define all four unconditionally. That was true of every
+ * Windows target this index had until a C library began presenting POSIX on
+ * one: over openkal the target is still PE, still Win64, and there is no
+ * `windows.h` --- `ares_setup.h:81` then reached for it and this repository's
+ * own compatibility measurement recorded
+ *
+ *     ares_setup.h:81:12: fatal error: 'windows.h' file not found
+ *
+ * `__has_include` is standard C and asks the question actually being asked.
+ * `HAVE_UNISTD_H` gets the same treatment for the same reason, from the other
+ * side: a POSIX-presenting Windows target HAS it. */
+#if __has_include(<unistd.h>)
+#define HAVE_UNISTD_H 1
+#endif
+
+#if __has_include(<windows.h>)
 #define HAVE_WINDOWS_H
+#endif
 
-/* Define to 1 if you have the winsock2.h header file. */
+#if __has_include(<winsock2.h>)
 #define HAVE_WINSOCK2_H
+#endif
 
-/* Define to 1 if you have the winsock.h header file. */
+#if __has_include(<winsock.h>)
 #define HAVE_WINSOCK_H
+#endif
 
 /* Define to 1 if you have the writev function. */
 /* #undef HAVE_WRITEV */
 
-/* Define to 1 if you have the ws2tcpip.h header file. */
+#if __has_include(<ws2tcpip.h>)
 #define HAVE_WS2TCPIP_H
+#endif
 
 /* Define if __system_property_get exists. */
 /* #undef HAVE___SYSTEM_PROPERTY_GET */

pkgs/c/compat.curl.lua

@@ -248,20 +248,56 @@ package = {
 #if defined(__linux__)
 
 #define CURL_OS "Linux"
-/* glibc's strerror_r returns char*, not int — curl needs to know which */
+
+/* `__linux__` ANSWERS "WHICH KERNEL", NOT "WHICH USERSPACE IS INSTALLED", and
+ * this block used to read it as the second. Every entry below that names a
+ * HEADER now asks whether that header is reachable, which is the question
+ * being asked and is standard C; every entry that names a GLIBC extension
+ * asks for glibc.
+ *
+ * Measured: over openkal-musl on x86_64-linux-gnu the kernel is Linux, so
+ * `__linux__` is correct, and `#define HAVE_LINUX_TCP_H 1` then sent
+ * `lib/setopt.c:31` to a uapi header that C library does not carry ---
+ * `fatal error: 'linux/tcp.h' file not found`, one of the failures in this
+ * repository's own compatibility measurement. */
+#if __has_include(<linux/tcp.h>)
+#define HAVE_LINUX_TCP_H 1
+#endif
+/* `__has_include` ANSWERS ABOUT THE HEADER, NOT ABOUT THE FUNCTION, and those
+ * two come apart here. openkal-musl CARRIES `sys/eventfd.h` and declares
+ * `eventfd` absent in its `[c-abi-absent]` table, so defining `HAVE_EVENTFD`
+ * from the header's presence produced a link that mcpp explained by name:
+ *
+ *     the C library in this graph (musl) declares that it does not supply
+ *     the following, and the link has just asked for it:
+ *       eventfd              as timerfd_create
+ *
+ * Neither is declared: curl has a poll-based path and takes it. A header
+ * test cannot stand in for a symbol test. */
+#if __has_include(<termio.h>)
+#define HAVE_TERMIO_H 1
+#endif
+#if __has_include(<sys/xattr.h>)
+#define HAVE_FSETXATTR 1
+#define HAVE_FSETXATTR_5 1
+#endif
+
+/* WHICH `strerror_r`, AND EXACTLY ONE OF THEM. glibc's returns `char*`,
+ * POSIX's returns `int`, and `strerr.c:30` refuses a configuration that names
+ * neither or both. Asking `__linux__` answered "glibc" for musl, which is the
+ * wrong one; leaving it out answered "neither", which is also refused. */
+#if defined(__GLIBC__)
 #define HAVE_GLIBC_STRERROR_R 1
+#else
+#define HAVE_POSIX_STRERROR_R 1
+#endif
+
 #define HAVE_GETHOSTBYNAME_R 1
 #define HAVE_GETHOSTBYNAME_R_6 1
 #define HAVE_ACCEPT4 1
 #define HAVE_PIPE2 1
-#define HAVE_EVENTFD 1
-#define HAVE_SYS_EVENTFD_H 1
 #define HAVE_SENDMMSG 1
 #define HAVE_MEMRCHR 1
-#define HAVE_FSETXATTR 1
-#define HAVE_FSETXATTR_5 1
-#define HAVE_LINUX_TCP_H 1
-#define HAVE_TERMIO_H 1
 #define HAVE_CLOCK_GETTIME_MONOTONIC_RAW 1
 /* Debian/Fedora layout; overridable at runtime with CURLOPT_CAINFO or the
  * SSL_CERT_FILE environment variable. */