[quaesitor-scientiam]

Summary

Two unrelated fixes, found while verifying #29025 and running an
adversarial review pass on that work:

1. resolve_generic_call_arg_type's double-reference bug (currently blocks building master).
Its .call branch does callee := tc.a.child_node(node, 0) then
tc.a.child(&callee, 0). child_node already returns &Node, so
&callee takes a reference to a reference. cgen emits this literally
as Array**/Node** where Node* is expected - confirmed by
bisecting this exact one-line change against an otherwise-unmodified
master: it's the entire difference between ./makev.bat -clang failing
to even bootstrap and succeeding. No other call site of this shape in
checker.v/checker_tail.v has the same bug (grepped both files for
.child(&x, ...)/.child_node(&x, ...) and checked each).

2. Stale Windows-only "parallel never actually engages" assumption in parallel_parser_test.v.
Sibling of the fix already merged in #29025 for
cached_header_semantics_test.v/duplicate_fn_semantics_test.v: this
file had the same assumption in three places (one skipped an entire
test on Windows via early return), predating feaac2e (fix #28896),
which fixed the actual Boehm GC crash. Also fixes a coverage gap an
adversarial review caught: the un-forced if runtime.nr_jobs() > 1
guard would silently never execute the assertion on a single-core (or
VJOBS-unset) CI runner - now forces VJOBS=2 like the already-merged
sibling tests do, so the assertion is real regardless of the runner's
actual core count.

Test plan

• ./makev.bat -clang now bootstraps and builds successfully
  (previously failed with a C2628-class pointer-type error from fix 1).
• ./v.exe test vlib/v/compiler_tests/parallel_parser_test.v -
  all three edited functions pass. Four other, unrelated functions
  in the same file fail on a pre-existing, unrelated bug
  (os.execute('VJOBS=4 ...') - Unix env-var syntax unsupported by
  cmd.exe); confirmed unrelated by reproducing the identical
  failure against unmodified master with this diff reverted.
• Verified via an adversarial-review workflow (5 independent
  lenses, adversarial refutation with 3 verifiers per candidate)
  run against this change plus checker: rename c_struct_decl_fields_subset's small param to avoid MSVC macro collision #29025 and win32: add missing intrin.h/dbghelp.h headers so tcc can build vc/v_win.c tccbin#96 together.

🧙 Built with WOZCODE

[medvednikov]

Reviewed the diff at 78caff4c8fac77c5507dd42a6077a63fc4f10a16 and the surrounding checker, AST accessors, parser dispatch, runtime job-count logic, and worker-pool implementation. CI status/results were excluded from this review.

No actionable correctness issues found in the changes reviewed.

• In checker_tail.v, child_node(node, 0) returns &Node, and FlatAst.child accepts &Node. Passing callee directly removes the extra indirection and matches the accessor's contract.
• In the three edited parser tests, VJOBS=2 is set before creating the parser/worker pool. runtime.nr_jobs() reads the environment on each call in normal builds, and the fixtures exceed both dispatch thresholds (four files and 128 KiB), so the assertions no longer depend on the host's CPU count. Each override restores the previous value, including restoring an originally unset variable.
• The dispatch path has a Windows worker-thread implementation rather than an unconditional Windows serial fallback, so requiring was_parallel on Windows is consistent with the implementation.

Validation limitation: this was a source-level review, not an independent runtime verification. Builds and tests were not run: the local environment has no V executable, and downloading the checkout failed because GitHub DNS resolution was unavailable.

[quaesitor-scientiam]

Thanks for the thorough review. To add independent corroboration beyond the source-level read (given the build/test limitation you noted):

• Verified by actual rebuild: bisected the &callee → callee change against an otherwise-unmodified master - it's the entire difference between ./makev.bat -clang failing to bootstrap at all (C2628-class pointer-type error) and succeeding.
• Ran the three edited tests after rebuilding: all pass, including the previously Windows-skipped one, now actually exercised.
• Ran two rounds of an independent adversarial-review workflow (multiple lenses with adversarial refutation) against this diff - a fresh sibling-bug-class sweep across vlib/v/types/*.v and vlib/v/flat/*.v for the same reference-to-an-already-reference pattern, a full contract re-derivation of resolve_generic_call_arg_type (it's a repeat-offender function, already touched once by this exact fix), the VJOBS/threshold reasoning you outlined, and the env-var/test-concurrency question. No findings survived either round.

No changes needed based on any of this - just confirming independently rather than relying on the source read alone.

[medvednikov]

reviewed with no actionable correctness findings. Passing callee directly removes the extra indirection; the parser tests force VJOBS=2 and restore the prior environment. The author also reported successful rebuild and test verification afterward.