Safety and traceability
Evidence (quality/run_evidence.sh, CI artifact evidence)
| File | Shows |
|---|---|
no_heap.md |
Every runtime header, every generated header of every model (examples, protocols and test models, among them an import of a MAVLink dialect) and profile, and the generated self-tests, compiled with -fno-exceptions -fno-rtti -DWIREFORGE_NO_STD_MUTEX at -O0 and -O2: the external symbols of each object. Anything outside a short allow list (compiler runtime, string functions; printing in self-tests) fails, so no object references operator new/delete, malloc/free, exception or RTTI support or threads. |
trace_<project>.md, .csv |
Every enum, checksum, type, message, field and interface: model source (file:line, the ICD row), C++ symbol, generated file:line, the high-level requirements (reqs) and derived low-level requirements (llrs) of the element, and the generated tests that verify them. A second table lists every low-level requirement with its parents and test count. Messages and interfaces without tests, LLRs without tests, LLRs verified by the runtime tests, requirements without model elements and derived LLRs are listed. |
icd_<project>.md |
Layout per profile: offsets, sizes, bit positions, frame sizes, link RAM. |
hashes.txt |
Model hash (sha256 of the definition content) of every project; the same hash is in the banner of every generated file, so generated code is tied to the exact model. |
Generated code carries the trace inline as well: each struct is documented with Defined in <file>:<line> (<JSON
pointer>) and each member with its file:line. In a project with a requirements catalog, every struct, enum,
checksum and interface also lists Requirements: <HLR ids>; <LLR ids>, and every generated test records the LLR ids it
verifies (RecordProperty("requirements", ...), docs/SCHEMA_SPEC.md section 17).
In no_heap.md, a header's object holds only what the header defines outside templates and uncalled inline functions,
so it shows that the header is clean; the self-tests instantiate the code, and the check fails when a message of an
interface is missing from them. The generated C (model.c and the interfaces) is checked the same way, compiled as C99.
The check covers the models of this repository. For the model of a user, the generated test project gives the same
answer: no_allocation_test.cpp counts the allocations of every message in C++, and in C the heap functions do not link.
A user type's own encode/decode is outside the no-heap check's object-file inspection unless its header is
compiled into an example the safety-evidence pipeline builds; this is the same "informative, not a conformance
claim" boundary as the rest of this table, not a new exception.
Independent expectations
The expected result of most generated tests is computed by the reference codec, which reads the model the same way as the code generator does. A misreading would show in both, so the tests could not catch it. Requirement vectors are bytes and values you write from the ICD; the reference codec only confirms them at validation time, and the generated tests compare with your values. That is the common-mode argument for keeping the two sources apart (docs/LIFECYCLE.md).
Coding-rule mapping
docs/CODING_RULES.md maps the concerns behind MISRA C++ and AUTOSAR C++14 to what enforces them here. It is a
coverage map, not a conformance claim.
Bare-metal targets define WIREFORGE_NO_STD_MUTEX (or WIREFORGE_DEFAULT_LOCK); the generated Handler and the
runtime ITransport have protected non-virtual destructors, so no deleting destructor (and no operator delete) is
emitted.