Stacked on #9431.

Motivation

SideEffectDetector::detect_side_effect_of_call_expr treats pure-annotated calls as fully side-effect-free and never recurses into the callee's body. So a pure-annotated IIFE like /* @__PURE__ */ (() => globalValue)() loses the fact that its body reads an unresolved global.

That matters because the two concepts are orthogonal:

Reading a top-level unresolved global IS an execution-order dependence even when the call is side-effect-free per the annotation. The current wrap trigger in impl_visit.rs:92-98 papers over this by using PureAnnotation as a "be cautious" proxy, which over-marks (row 1 wraps too) while under-modeling (the actual signal — GlobalVarAccess from the body — is missing).

Changes

• Add iife_body() free function: returns the FunctionBody if callee.get_inner_expression() is ArrowFunctionExpression or FunctionExpression. Transparent wrappers ((callee), callee as T, etc.) are peeled by get_inner_expression.
• In detect_side_effect_of_call_expr, when the call is pure-annotated and iife_body() returns Some, walk each body statement through the existing detector and OR in only the GlobalVarAccess bit. Unknown is deliberately masked off — the annotation already certified no side effect at the call level; this scan exists strictly to harvest the order-sensitivity signal.

Why no snapshot churn

Purely additive. The wrap trigger still intersects with PureAnnotation, so cases that previously wrapped continue to wrap. The new GlobalVarAccess signal sits alongside PureAnnotation for now.

Phase 1 of a multi-step reconcile

This PR is the first of three steps to resolve a design tension in SideEffectDetail::PureAnnotation:

1. This PR — produce GlobalVarAccess for pure-annotated IIFE bodies. Additive. No behavior change.
2. Follow-up — drop PureAnnotation from the wrap trigger in impl_visit.rs:92-98. Reconciles with the sibling branch fix/remove-pure-annotation-execution-order-sensitive. Modules that were over-wrapping (pure call to local empty function with no globals) stop wrapping; modules that need wrapping (pure IIFE reading globals) keep wrapping via the GlobalVarAccess path this PR adds. Will have snapshot churn — that's the point.
3. Cleanup — drop the SideEffectDetail::PureAnnotation flag entirely if no consumer remains.

Not addressed

Parameter defaults that evaluate globals at call time — /* @__PURE__ */ ((x = globalValue) => 1)(). The x = globalValue runs but isn't inside body.statements. Noted as a follow-up in the code comment. Probably a small extension to iife_body()-adjacent code; not blocking the Phase 1/2 reconcile.

Tests

• test_pure_annotation_propagates_through_compound_expr — existing test, expectations updated. Cases that read globalValue now expect PureAnnotation | GlobalVarAccess (the new combined semantic). Cases with bodies that don't read globals ((() => 2)(), (() => true)(), (() => 'k')()) still expect just PureAnnotation.
• test_pure_iife_body_global_propagates_global_var_access — new test, direct coverage for the body scan: arrow with expression body, arrow with block + explicit return, function expression IIFE, pure IIFE with no globals (negative case), member expression rooted at an unresolved global.

cargo test --package rolldown --lib ast_scanner::side_effect_detector::test::
# 37 passed

cargo test --package rolldown --test integration -- strict_execution_order tree_shaking no_side_effects
# 102 passed, 0 failed