2026-09-02 - (/loop) Bulk port batch 2: the runtime's out-of-line atomics, conversions and assert

Second batch under D472, and the first where the honest answer to part of the list was "not yet, and here is why".

documented   711 needed, 441 missing   ->   711 needed, 434 missing

What went in - seven functions

The C11 atomics (new orbistoun-libc/src/atomic.rs): _Atomic_load_4, _Atomic_fetch_add_4, _Atomic_fetch_sub_4, _Atomic_compare_exchange_weak_4. <stdatomic.h> is mostly compiler intrinsics, but a runtime that must work without them ships out-of-line copies and calls those; the Dinkumware runtime this platform carries (D468) names them this way.

Two decisions worth stating:

• The memory-order argument is deliberately ignored, upward. Each takes one, and the strongest order is a conforming implementation of every weaker one - SeqCst never permits a reordering Relaxed forbids - so all four use SeqCst rather than mapping an enumeration whose numbering is the runtime's own and is not established here. A deliberate strengthening costs a little speed and cannot give a wrong answer; guessing the enumeration could.
• A misaligned object is refused rather than dereferenced. C11 requires atomic objects to be aligned (6.2.8), so a misaligned address is a guest already in undefined territory - but forming a misaligned reference would make it undefined here too, and unlike the guest's mistake ours would be silent.

Conversions: _Stoul, _Stoull - the runtime's own names for what strtoul/strtoull wrap. Delegated to the existing implementations rather than reimplemented: two copies of one parser is two answers to one question, and the copy nobody exercises is the one that drifts.

_Assert - the runtime routes a failed assert here instead of calling abort itself, so this is where a guest says which assertion it lost. Message to stderr and the kernel log first, then the ordinary abort stop; stopping without the message would throw away the only part a reader needs (ISO/IEC 9899 7.2.1.1).

Six tests, each guard made to fail with its passing case beside it. The two that matter: fetch_add answering the previous value (answering the sum passes every single-threaded eyeball test and breaks every ticket lock), and a failed compare-exchange writing the actual value back through expected - without which the caller's retry loop never terminates.

What was deliberately left out, and why

• _FCosh, _FSinh, _FExp, _FDtest, _FNan. These are not the C library's coshf and friends: they are the runtime's internal helpers with their own argument conventions, and _FDtest answers a classification using the runtime's own constants, which are not established here. Writing them would mean inventing those numbers, which is the one thing principle 3 forbids outright. _FNan is a data object rather than a function besides.
• _Lockfilelock, _Unlockfilelock, _Locksyslock, _Unlocksyslock. Implementable, but correctly: they are recursive locks, and the tempting version - a no-op, since much of a guest is single-threaded early on - is a silent correctness hazard the moment two threads share a FILE. They need a real recursive lock keyed by the argument, which is a batch of its own rather than four one-line stubs.
• _Thrd_id, _Thrd_join. These belong beside the _Mtx_*/_Cnd_* family in orbistoun-kernel, not in libc, and want the thread registry rather than a shim here.

_Stdout / _Stderr - not a new design question after all

Checked before writing, as the plan said to. A data-import mechanism already exists (D323): orbistoun-thunk reserves storage per named data import and publishes the addresses, and data_symbol(name) reads one. So these need no new concept - they need the right contents: a FILE * the rest of libc recognises. fprintf already routes by wrapped descriptor (orbistoun_fs::open::wrapped_descriptor) and falls back to the host's error stream otherwise, and wrap_descriptor(fd) mints exactly the handle required. So the work is to wrap descriptors 1 and 2 at start-up and store those handles in the two objects - real work, but plumbing rather than a decision. No D473 needed.

clippy --tests clean, fmt clean, orbistoun-libc tests pass, identity scan clean, nothing committed.

Next: _Stdout/_Stderr by the route above, the four recursive locks done properly, then batch 3 - the Itanium C++ ABI (_Unwind_Resume, the _Z* mangled set).