test(gc): temp-root codegen contract lives only in crates/perry-codegen/tests/, which does not run per-PR

[proggeramlug]

Split out of #6983, where it was raised in review and explicitly declined rather
than silently skipped.

The gap

All of the temp_root family's codegen coverage lives in
crates/perry-codegen/tests/ — temp_root_argument_temporaries.rs (#6972) and
temp_root_operand_temporaries.rs (#6983), 10 tests between them. Per CLAUDE.md:

Integration suites under crates/*/tests/*.rs do not run per-PR (nightly/tag
only) — a regression there can land green and sit red for days. Prefer putting
acceptance coverage in cargo-test-visible unit tests (#5960).

So the assertions pinning the rooting emission contract — push → re-read →
consuming call → release, and the interleaving requirement — are CI-invisible on
the PRs most likely to break them. The runtime half (gc::tests::temp_roots,
in src/) does run per-PR; the codegen half does not.

That matters more than usual for this family, because the failure mode is
silent. A lowering path that quietly goes back to threading a value through an
SSA register produces correct output under the default configuration — the
conservative native-stack scan pins it by accident — and only diverges under
PERRY_CONSERVATIVE_STACK_SCAN=off or an evacuating minor. The emission-contract
tests are the fast signal; the corpus arms are the slow one.

Why it was not fixed in #6983

The assertions need a built Module plus compile_module, so moving them into
src/ unit tests means either constructing a FnCtx by hand or duplicating the
~90 lines of entry_opts() / module_with_init() harness. Doing that once for
the whole family is worth it; doing it piecemeal for one PR's three tests is not,
and would leave the two files inconsistent.

Suggested shape

Extract the harness (entry_opts, module_with_init, ir_for, plus the
ir_for_new class builder from #6983) into something both a src/ unit-test
module and the existing integration tests can call — a #[cfg(test)] helper
module in perry-codegen, or a small pub(crate) test-support module — then
move at least the ordering/interleaving assertions across:

• argument accumulator: push → fused push → re-read → truncate (fix(gc): make argument temporaries precise roots (#6951) #6972)
• MapSet: receiver rooted before the key is lowered (fix(gc): operand temporaries are precise roots — collection methods, constructor args, string-method receiver (#6970, #6969, #6971) #6983)
• constructor args: push(arg0) before argument 1's own allocation (fix(gc): operand temporaries are precise roots — collection methods, constructor args, string-method receiver (#6970, #6969, #6971) #6983)
• the negative gates: shapes that must emit no rooting calls at all

The negative gates matter as much as the positive ones — they are what keeps the
fix from silently becoming a cost on hot paths.

Related: #6951, #6972, #6975, #6983, #5960.

[jdalton]

Heads-up on a small campaign I'm lining up around the rooting family, routed through this issue first — flagging it here so nobody double-works it and so the sequencing is easy to veto.

The plan, in order:

1. test(compat): pinned-Node builtin-module compatibility matrix #7074 (Node builtin compat matrix) — rebased onto current main today and mergeable again; independent measurement infra, not part of the stack.
2. test(gc): assert the child relocation, not a whole-process copied count #7092 (just opened) — fixes gc test: large_object_old_born_array_slot_write_keeps_young_child_alive asserts an exact whole-nursery copied_objects count and fails on macOS (4249 vs 1) — pre-existing on main #7060 so cargo test -p perry-runtime is green on macOS while working in the GC area. Also independent.
3. This issue (test(gc): temp-root codegen contract lives only in crates/perry-codegen/tests/, which does not run per-PR #6988) — extract the temp-root harness so the rooting-contract assertions run in per-PR cargo test rather than nightly-only crates/perry-codegen/tests/. This becomes the base of a stack.
4. Stacked on it, in dependency order: gc: constructor arguments on lower_new's non-class branches (Readline / imported ctor / new Function) and builtin.rs are still not precise roots #6986 (constructor-argument rooting on lower_new's non-class branches), gc: a closure's captured this is stored in an alloca that is not shadow-bound (methods bind theirs) — same family as #7055 #7066 (closure-captured this shadow binding), and runtime: a fourth unrooted-operand family — js_string_coerce as a plain ToString argument coercion (string built-ins, constructors) #6949 (the js_string_coerce ToString-coercion rooting family) — each carrying a sabotage-verified regression test that lands in the per-PR-visible harness from step 3.

How the stack works from a fork

I contribute from a fork and can't create base branches in this repo, so this is an annotated stack rather than a GitHub-native one: each PR targets main, is opened in dependency order with a "stacked on #X" note, and until the layer below merges its diff will include that layer's commits. As each layer squash-merges I'll rebase --onto the layers above so the diffs collapse to just their own change. If any of these is already spoken for, or you'd rather the harness extraction wait, say the word and I'll reorder.