libc: sinhf and asinl are absent, first stops of Ruffle and RetroArch on main; bit-exact ports ready #3519

libc (libSceLibcInternal) does not export sinhf (NID 1t1-JoZ0sZQ) or asinl (NID Nx-F5v0-qU8). On current main 0f863c10 (fresh Windows relink, --windows --registry unused-filter=0) Ruffle Flash (PPSA68091) stops on sinhf and RetroArch (PPSA99169) on asinl:

FAIL: unresolved ELF import 1t1-JoZ0sZQ from libSceLibcInternal.prx
FAIL: unresolved ELF import Nx-F5v0-qU8 from libSceLibcInternal.prx

Fix branch ready: screamm:feat/libc-sinhf-asinl (two commits on main 0f863c10: 80ee7d83 feat(libc): implement sinhf, 62c4c7c3 feat(libc): implement asinl). I cannot open it as a PR because of the intake check; anyone eligible is welcome to carry it, keeping the commits as they are. The closed #1702, #2037, #1405 and #2271 were host one-liners; this is different.

• Both are bit-exact against the console's own code: the libc.prx bundled with PPSA28319 holds real code for sinhf (0x51d70) and asinl (0x429f0) under these NIDs; it was disassembled and run on this machine through a harness and compared with the ports. The console code is FreeBSD msun plus a Sony errno helper.
• sinhf: a port of the older msun e_sinhf.c (FreeBSD 5aa554c7) with the expm1f/expf paths it reaches from stable/9 (59bc2f09); a host-expf version differed from the console on 4% of inputs by up to 4 ULP, the port matches on all 2,238,032,378 finite inputs up to the overflow threshold; above it errno ERANGE and ±inf with FE_OVERFLOW; NaN and ±inf return x + x, ±0 keeps its sign.
• asinl: a port of e_asinl.c with the ld80 invtrig.c coefficients (9f043204), whose 80-bit constants the console binary holds; matches on 20 million random arguments in (-1, 1) plus the branch edges; ±1 gives ±pi/2 rounded to 64 bits; |x| > 1, ±inf and NaN raise FE_INVALID, set errno EDOM (NaN too, a Sony quirk) and return NaN. Nothing throws.
• guest_math pins exact bits for every branch, sign and boundary plus the errno and floating-point flag behaviour. TechnicalDebt gets one bullet per function, including a finding beyond this change: on Windows a guest thread started with _beginthreadex runs the x87 unit at 53-bit precision, so every long double export is less precise in guest threads than on FreeBSD (64-bit); that belongs in guest thread creation.

Tested on Windows 11 Pro 26220, WinLibs MinGW-w64 GCC 15.2.0 r7, Ninja, Release, BUILD_TESTING=ON, full build including libs: guest_math 3/3; ctest -j8 -E "^agc_driver_|^video_out_flip" 533/534, the one failure being windows_exceptions (#3411); with main's Math.cpp the extended test fails; tools/check_conventions.py --base main and git diff --check pass. Not run on Linux, no title run.

Written with Claude Code assistance (implementation by a subagent, reviewed and validated in this session).