[SizzleUnrlsd]

Item 2 of #21. Stacked on #22 (test/api-consumer): merge that one first.

Change

1. Every fixture states what it proves. The 43 sources under test/ carry a header, one line per expectation, read by test/check_fixtures.py (the fixtures test):

   // args: --ct-modules=alloc,vtable --ct-vtable-diag
   // expect: rule=vtable-null-this cwe=CWE-476 level=error line=11
   // expect: none
   // expect: program-exit=3
   // expect: output="auto-free ptr="

   Each fixture runs with its arguments (--ct-modules=alloc,bounds by default); the SARIF results must match the declared findings one to one, faulting line and allocation site in the fixture as compiled; the analyzer's exit status must be 1 with findings and 0 without. program-exit and output check the program itself. expect[linux]: and expect[darwin]: keep a line to one platform. A fixture without an expectation fails, which is what the checker did on its first run, on all 43.

2. What the fixtures now prove. The 7 memory and 5 vtable-diagnostic fixtures keep the expectations of the old table. The 14 autofree fixtures get back the expectations of run_autofree_tests.sh: auto-freed allocations leave no leak and print the auto-free line, the three escaping ones leak and exit 1. The five other vtable fixtures run with the vtable module and report nothing; ct_vtable_uaf.cpp is reported by the alloc module, as the access faults before the vtable module sees the freed vptr (with the vtable module alone it segfaults with no finding). ct_realloc_zero.c and ct_autofree_sbrk.c differ by platform: realloc(p, 0) frees on glibc and leaks on macOS; sbrk is a real allocation on Linux only. The vtable-diagnostic fixtures' hand-written site strings now name the line of the call itself (they pointed two lines above it).

3. test/cli/check_cli.py (the cli test) keeps the CLI checks of the old findings script: text output, the SARIF log of a batch, exit 2 for what cannot be analyzed.

4. Deleted: check_findings.py (replaced), sweep_compiles (sweep_runs builds the same tree and keeps the batch heuristics, -x c++ and -DDEBUG, covered), and the template leftovers registered nowhere: test/examples/test_*.py, run_autofree_tests.sh, test/scripts/, test/docker/.

5. The header lines shift heap_overflow_write.c by one line: check-analysis.sh and api_test.cpp follow. README: the tests table and how to write a fixture.

Tests (written first)

check_fixtures.py ran before any fixture had a header: 43 failures, "states no expectation". With the headers, 43 pass on macOS. With ParseFindings disabled in Run: 40 failures in fixtures and cli fails too.

Verification

• macOS arm64: ctest --test-dir build 14/14 (fixtures 16 s, cli 4 s); ./scripts/format-check.sh clean.
• Linux: this pull request's Build workflow. The two platform-specific expectations are predictions for Linux; the workflow decides them.