Cut test suite runtime in binary roundtrips and integer sweeps

[ayush-singh-0601]

See #5418.

Linux CI runs --no-skip, so the expensive binary roundtrip cases actually run there. Each corpus file was parsed four times (once per input adapter). Parse once and reuse it.

The 16-bit integer sweeps still prove width selection: stride 7 hits every low-byte residue, and the endpoints stay in the loop.

unit-large_json used depth 5,000,000 to check the non-recursive destructor from #1419. 500,000 is enough. jeopardy.json is ~500 MB of serialization output, so it is now its own skipped test; canada/twitter/citm/sample keep running.

Unicode byte sweeps are left to #5426.

• The changes are described in detail, both the what and why.
• If applicable, an existing issue is referenced.
• The Code coverage remained at 100%. A test case for every new line of code.
• If applicable, the documentation is updated.
• The source code is amalgamated by running make amalgamate.

[nlohmann]

Please rebase to the latest develop branch as I had to fix some tests.

[ayush-singh-0601]

Rebased onto the latest develop.

The gcc job was failing as well, so I fixed that in the same push. static_cast<int32_t> on the sampled loop bounds is a useless cast there (int32_t is just int), and ci_test_gcc builds with -Werror. Those bounds are plain ints now. The ranges, the stride of 7, and the endpoints are unchanged.

[nlohmann]

Thanks! Most of this PR is fine to merge. The stride-7 integer sweeps still hit every boundary and every byte value, the roundtrip hoisting only removes duplicated parsing, and 500k depth is still plenty for unit-large_json.

I'd like one change: please revert tests/src/unit-binary_formats.cpp to develop's version.

The split doesn't save time, because of how skipped tests run in CI (cmake/test.cmake):

Job type	Command	Runs skip() tests?
Default (gcc/clang/sanitizer/coverage)	test-foo --no-skip	yes
JSON_FastTests=ON	test-foo	no
Valgrind	valgrind test-foo	no

On develop, "Binary Formats" is skipped, so it runs only in the default jobs. With this PR:

• Default --no-skip jobs still run both test cases, jeopardy.json included, so there is no saving.
• FastTests and valgrind jobs never ran any of this before. Now they run canada/twitter/citm/sample. Under valgrind, test-binary_formats_valgrind went from about 2 s to about 223 s (compared with develop CI run 35829411620). That makes the valgrind job slower overall (ctest time 390 s → 527 s), even though the other changes should have made it faster.

So the cheaper files don't "keep running": they never ran in those jobs. It isn't a coverage gain either. The only checks that run in more places are the encoded-size checks for four files, and the default and coverage jobs already run those.

If you'd like something from that file in the fast jobs, un-skipping just sample.json would be acceptable, since it's small. Otherwise, a plain revert is best.

Nit: the comment on next_integer_sample in tests/src/test_utils.hpp says "Walk [first, last]", but the function has no first parameter.

---

This review was written by Claude Code on behalf of @nlohmann.

[ayush-singh-0601]

Addressed both points from the review:

1. Reverted tests/src/unit-binary_formats.cpp to develop. The split was counterproductive -- the cheaper corpus files never ran in fast/valgrind jobs before, and introducing them there inflated valgrind runtime from ~390 s to ~527 s with no coverage gain. Everything is back to a single skipped test case.

2. Fixed the next_integer_sample comment in tests/src/test_utils.hpp. The old wording said "Walk [first, last]" but the function has no first parameter. Updated to "Advance i toward last (inclusive) by stride", which matches the actual signature.

No other files were touched. Ready for another look when you have a chance.

[nlohmann]

Looks good to me.