2026-08-19 - Depth: 27 positive checks across libc and maths

Done. Added behavioural depth where signatures are certain, rather than more census breadth.

• 035-libc extended from 10 to 20 checks: strcpy/strcat, strncpy padding, strstr, memchr/strrchr, strspn/strcspn, strtok, bsearch, ctype, seeded rand, strtoul base detection.
• New 037-math: 7 exact-value checks. No tolerances anywhere - every value is exactly representable, because an epsilon is where a wrong answer hides.
• 24 names promoted out of the census into real declarations.
• docs/BACKLOG.md gained item 4b (setjmp/longjmp, deliberately absent).

Verified. 14 sections, 77 checks. 035-libc 20/20 and 037-math 7/7 against real glibc. 400 imports, report well-formed.

Surprises.

• A weak undefined reference does not make the linker search libraries. Adding -lm changed nothing: no strong reference to libm existed, so --as-needed (the default) dropped the dependency and every maths symbol stayed null. --no-as-needed forces the DT_NEEDED entry, after which the weak references bind at load. The distinction is visible in the symbol table - resolved imports read FUNC WEAK UND strlen@GLIBC_2.2.5, unresolved ones read a bare NOTYPE WEAK UND with no version and no type.

• The first failure to fix was mine, and make -n found it in one line. The recipe never referenced $(HOST_LIBS); an earlier edit had defined the variable and not wired it in, so two rounds of linker theorising were spent on a build that was never passing the flag at all. Printing the command should have been the first move, not the fourth.

• Seven skipped checks were the harness working, not failing. Before libm was linked, 037-math skipped every check with "the symbol is not present on this platform" rather than jumping to null and taking the process down. That is exactly what the weak-symbol presence check was added for, demonstrated by accident.

Not done. Everything blocked stays blocked: struct-taking functions, NID verification, and running under an emulator at all. The positive/negative ratio has moved a long way but only inside two sections - the vendor subsystems are still checked almost entirely from the failure side.