nx_ccbuild_lib.nx

nx_ccbuild_lib.nx -- LIB: COMPILE A SOURCE WITH A NAMED COMPILER, THEN RUN IT. WHY THIS EXISTS. Every compiler gate needs the same four moves: fork/exec a chosen cc, assemble+link its output sovereignly, make the product executable, and run it capturing status. nx_cc_equiv_gate had the only correct copy, and the next gate that needed it (nx_linecont_gate, LN15) would have COPIED it -- which is the duplicate-ruler defect: two spellings of one invariant that drift silently. nx_spendgate/nx_capsearch over 6,925 organs (corpus_complete=1) returned NO incumbent exposing this primitive, so it is extracted here once. The bodies below are VERBATIM from nx_cc_equiv_gate's eq_* helpers so the migration of that gate onto this lib is provably neutral: its oracle is its own published verdict, rows=10 passed=10 selfhost=1 -- anything else means the extraction changed behaviour and must be reverted, not argued with. EVERY CORRECTION THAT LIVED IN THOSE HELPERS IS CARRIED, NOT RE-LEARNED: * nxasm writes 0644, so the product is NOT EXECUTABLE and every exec dies 127. Both sides dying identically made two failures agree and READ AS AN EQUIVALENCE. The fchmodat is load-bearing; do not drop it. * a failed rename used to be DISCARDED, so elf_out went missing while the build still returned 0 and the downstream 127 looked like a compiler fault. * the compiler's stderr must go to a REAL log, never /dev/null: a gate that reports a failure without its diagnostic makes the failure unactionable. * every source path is relative to the buildroot, so a caller with a different cwd (nx_job_run fixes cwd at ~/nishihost) fails to FIND its sources and every row reports build-failed -- which reads as a broken compiler when nothing is wrong. cb_anchor_root exists to make that impossible; call it FIRST. ONE DELIBERATE CHANGE FROM THE ORIGINAL: eq_build carried an `o_tmp` parameter its body never read. It is dropped here rather than carried forward, and the neutrality of dropping it is exactly what the equiv gate's re-run proves.