LLM-breakdown of 3 issues when run through nighlies/mctest #389

@g5t since ~ last ADR discussion over at McCode, the nightlies and the testsuites report 3 errors for the mcstas-antlr run with 3 new LLM-built instruments:

• Test_NeXus_comp_par_runtime gives a runtime error
• Unittest_INHERIT_DIAPHRAGM is not compiled
• Unittest_JUMP_RELATIVE is not compiled

I put my LLM to work on the outputs and got the below analysis:

All three failures are on the mccode-antlr side, and none comes from our patches. I tested with mccode-antlr 0.33.0, the current PyPI release, against current McCode main. It generates code with the 3.9.2 release runtime, which is what your CI’s antlr setup ends up using. I didn’t touch your checkout.

1. Unittest_JUMP_RELATIVE: the code generator crashes (reproduced).
   The crash is in mccode_antlr/instr/visitor.py:

def visitJumpPrevious(self, ctx):
    i = ctx.IntegerLiteral()
    return ("PREVIOUS", -1) if i is None else (f"PREVIOUS_{i}", -int(i))   # TypeError

ctx.IntegerLiteral() returns a parse-tree token, not a number, so int(i) fails. visitJumpNext has the same bug. Any JUMP PREVIOUS(n) or JUMP NEXT(n) with an explicit n crashes, while plain JUMP PREVIOUS / NEXT works. The fix is int(i.getText()) in both functions. The later resolution (absolute_target = index + relative_target) is already correct. Separately, there’s no range check, so an out-of-range target isn’t reported as an error, unlike classic McStas after the grammar fixes.

2. Unittest_INHERIT_DIAPHRAGM: the C doesn’t compile (reproduced).
   The error is redefinition of 'slit_print_if' (and the other slit_* helpers). component_share_declarations() in translators/c_decls.py writes out each component type’s SHARE blocks with no deduplication. Diaphragm inherits Slit’s SHARE, so it’s written once for Slit and again for Diaphragm. Classic McStas emits each SHARE block once, as the grammar-fixes ADR (ADR_20261007_GRAMMAR_FIXES.md) requires. Everything else about the INHERIT works: with the duplicate block deleted by hand, it compiles and gives AfterSlit_I = 0.2497 and AfterDiaphragm_I = 0.0621 (expected 0.25 / 0.0625). The fix is to skip SHARE blocks already emitted, matched by content or source file. The loop also iterates over a set, so the SHARE order isn’t deterministic.

3. Test_NeXus_comp_par_runtime: runtime error (from reading the code; I couldn’t build with NeXus here).
   mccode-antlr and classic McStas describe component parameters to the runtime differently. The runtime function mcrt_register_param() rebuilds each parameter struct’s layout from the type name passed in each mccomp_param_nexus(...) call. It only understands classic McStas’s names: int, char*, char, MCNUM, double, plus a {...} literal for vectors. mccode-antlr passes something else:

As a result psd, psd_copy, lmon, banana, both benders and conics get no runtime_value. dump_runtime.py then reports failed expectations, and the instrument’s FINALLY calls exit(1), which the test harness reports as a runtime error. I expect the downloaded logs to show unsupported:unknown-type for those components in the h5 file.

This one can be fixed on either side:

mccode-antlr side: emit classic’s type names and literals. The struct also has to store {...} vectors as arrays, otherwise the vector values still can’t be read.
McCode runtime side: have mcrt_register_param() also accept "char " and "double". If it should report the bender vector values, the element count can come from the comma-separated literal and it can read through the pointer.

The first two are clear-cut mccode-antlr bugs; I’d file them on mccode-dev/mccode-antlr.