D290 - Every floating-point function works and every one is reported as missing

decided · 2026-08-26 · a first explanation that was wrong, and the real one underneath

One run of the conformance probe says both of these:

• 037-math/sqrt passes, against spec, along with twelve more maths checks.
• The run report says libc::sqrt, round, pow and trunc were called and nothing implements them.

The first explanation was sceKernelDlsym - unimplemented, called seven times - on the reasoning that a symbol resolved at run time by name arrives somewhere different from the same symbol in the import table. That was wrong, and the probe's own transcript says so: 110-modules/symbol is skipped, with the reason "no module loaded, so there is no handle to ask". The probe never resolved anything through dlsym, so nothing reaches maths that way. A mechanism was asserted that had not been established, which is the failure principle 3 names, committed while writing the entry that names it.

The real cause is accounting. orbistoun-thunk keeps two handler tables - one for functions answering in rax, one for functions answering in xmm0 (D268) - and is_implemented reads only the first. So every floating-point function dispatches correctly, computes the right answer, passes its conformance check, and is recorded as a call nothing implemented.

Three things follow from it, and the third is why this is a decision rather than a bug fix:

• Argument dumps are taken for it. Dumps fire only for imports nothing implements, so every maths call has been dumping six integer registers that hold leftovers - a float function takes nothing in them.
• The findings list is wrong, and it ranks by call count, so the loop has been proposing work on functions that were finished.
• standing is understated. That is the number PROJECT_STATUS.md says to read rather than the call count - "the share of calls answered by a real implementation rather than a placeholder" - and it is the project's headline measure of its own progress.

The comment that got this right is one layer up, in symbols::all(): "Both tables. A function that answers in xmm0 is as implemented as one that answers in rax, and counting only the first would report the maths library as missing while it worked." Somebody wrote that sentence, applied it there, and the layer underneath still asks the shorter question.