jit: daslang -lib emits a C-ABI library

[aleksisch]

A host that only needs to call one daslang script had two options, and both asked
too much. The C++ API embeds the compiler. Standalone AOT emits C++ source the host
has to compile itself, and the result is a C++ class. Neither works when the host
is C, or when it wants a plain ABI boundary.

daslang -lib script.das -output build/script now writes a native library and the
C header a host calls it through. --jit-lib-static gives an archive instead. Which
functions cross is your choice: [export_c] marks them one at a time, and
-lib-export-all offers every public function of the entry module whose signature C
can spell, naming the ones it skips. Selection reads one bit, Function.flags.exports,
so the interpreter, AOT and the header emitter cannot drift apart. Scalars, string,
pointers and enums cross by value; structures and the float2..uint4 / range
families cross as const T * and return through a trailing out pointer. Every
declared structure carries a size assert and one offset assert per field, so a host
built for another target fails to compile rather than misreading memory.

A library is called by code that cannot catch anything, which shapes the entry
points. Every thunk runs its body under the context's catch boundary: a das panic
returns zero, leaves the out param untouched, and reports through
<P>_last_error(ctx). <P>_create works on any thread, because daslang's
environment is thread-local while the registration behind it happens once. A second
daslang runtime in one process declines with a null instance instead of aborting.
And <P>_destroy emits the program's [finalize] calls itself, because a jitted
SimFunction carries no shutdown flag for the runtime's own drain to find.

Where to look: modules/dasLLVM/daslib/llvm_lib.das is the entry and thunk emitter,
daslib/c_api_header.das decides what can cross and writes the header, and the four
emit_standalone_* helpers in llvm_exe.das are the exe startup both artifacts now
share. The riskiest spot is the thunk ABI: a structure result is built in an alloca
of its own alignment, never in a frame field, because an in-struct [N x i8] slot
aligns to 1 while the body stores through it at the structure's real alignment.

Validation, claims, ledger

Validation

• untracked gate is red on 6388 files that predate this branch and belong to the
  working copy, not the change: four build directories, a soundfont, model weights and
  scratch scripts. Nothing from this branch is untracked.
• tests-interp reports 13102/13113. Both failures are wall-clock budget overruns,
  not assertions: tests/lint/test_stale_nolint.das and
  tests/lint/test_parallel_equivalence.das against --max-file-time 60. Measured
  cleanly, twice each, they are 46.4s / 46.7s and 42.2s / 42.4s; the box was at load
  12.7 on 32 cores during the run. Neither test reaches code this branch changes.
• sync gate could not run: it fetches over https and GitHub was rate-limiting.
  Fetched over ssh instead and rebased onto origin/master at c409fcc2a; the
  upstream diff and this one share no file.
• Local-only, beyond what CI runs: a C host driving both artifacts through the
  generated header - scalars, struct in and out, float3 in and out, string round
  trip, bool, int64/uint64, two independent instances, an instance created on a second
  thread, a panic returning zero with its text readable and the instance still usable
  afterwards, an out param untouched after a panic, and [finalize] on destroy. The
  same host links the static archive. A C++ TU compiles the header on its
  static_assert branch, and perturbing a structure makes the layout asserts fire.
• The alignment fix is negative-controlled: test_jit_lib_result_slot_is_aligned
  fails when the result slot moves back into the argument frame.

Claims - stated, not tested

• Windows and macOS take untested branches: the .dll / .dylib artifact names, the
  MSVC import-library link in the ctest driver, and llvm-lib / lib as the archiver.
  Verified by reading against the neighbouring create_shared_library, which the exe
  path already exercises on those platforms. A break would be a link failure in the
  lib_capi ctest, not a wrong answer at run time.
• --jit-lib-static reuses --jit-path-to-linker to name the archiver. Recorded in
  that flag's help text rather than given a flag of its own.

Not done

• One daslang library per process. jit_register_Module_* carries no already-created
  guard, so a second library and an embedding host both decline. Lifting it means
  guarding the registration, which is code the exe path shares.
• No cross-compilation: -lib with --jit-target fails early.
• The full test_aot sweep did not run; test_aot_subset, what PR CI builds, does.
• 81 sites across daslib/, modules/dasLLVM/daslib/ and utils/lint/ hand-roll the
  annotation-name scan that a has_annotation(fn, name) helper in ast_boost would
  cover. Two of this branch's dupe hits are instances. Left alone as a tree-wide
  refactor, surfaced here as a candidate.

🤖 Generated with Claude Code