…10351)

### Description

Fixes #10337.

Since 1.1.5, bundling a dynamically imported `sideEffects: false` package (three-mesh-bvh in the report) together with a `codeSplitting` group that matches the package crashes with:

```
Symbol "common_functions" in ".../three-mesh-bvh/src/webgl/glsl/common_functions.glsl.js" should belong to a chunk
```

### Root cause

Regression from #10080. When a manual code-splitting group leaves a dynamic entry's facade chunk empty, `optimize_facade_entry_chunks` eliminates the facade and re-includes the entry's namespace object (`SymbolIncludeReason::SimulatedFacadeChunk`) so the namespace becomes an export of the merge-target chunk.

That include went through `drain_body_demand_stmts`, so it counted as *body demand* for the entry module and resurrected its deferred (gated) side-effect statements. In the reproduction, three-mesh-bvh's backwards-compatibility statement

```js
export const shaderIntersectFunction = ` ... ${BVHShaderGLSL.common_functions} ... `;
```

comes back to life, and its member expression canonicalizes through the `export *` chain straight to the glsl leaf module. The leaf therefore becomes included **after** `split_chunks` already assigned every included module to a chunk, so it belongs to no chunk and `compute_cross_chunk_links` panics.

### Fix

A simulated-facade include only materializes the eliminated facade's namespace object as a chunk export; it is not a semantic observation of the module, so it must not create body demand. `handle_include_symbol` now skips `drain_body_demand_stmts` for `SimulatedFacadeChunk` includes — mirroring the existing `WorkItem::Module` gate on `is_simulated_facade_chunk` a few lines below. This is safe because the namespace statement's symbol getters were already narrowed to link-time-used symbols before the replay, so the include cannot legitimately need anything that wasn't included at link time.

This also restores the invariant that chunk-layout choices don't change tree-shaking results: with the fix, the grouped build emits exactly what the ungrouped build emits (the dead `${common_functions}` statement no longer appears).

### Alternatives considered

Instead of (or in addition to) the semantic gate, the generate stage could detect modules newly included by the replay and assign them to the merge-target chunk. That converts the panic into working output, but it ships code the ungrouped build proves dead, and it papers over inclusion growing after chunk assignment instead of preventing it — so the semantic gate alone is preferred.

### Test

`crates/rolldown/tests/rolldown/issues/10337` mirrors the three-mesh-bvh shape: dynamic `import()` → an index module with `export * as` plus the deferred template-literal statement → an `export *` barrel → the glsl leaf, with `treeshake.moduleSideEffects: false` and a `codeSplitting` group. The entry only touches `res.MeshBVH`, so the leaf is reachable exclusively through the facade-elimination replay. The fixture panics on `main` and passes with this fix; it also executes the output under node and asserts the namespace member exists.

…10351)

### Description

Fixes #10337.

Since 1.1.5, bundling a dynamically imported `sideEffects: false` package (three-mesh-bvh in the report) together with a `codeSplitting` group that matches the package crashes with:

```
Symbol "common_functions" in ".../three-mesh-bvh/src/webgl/glsl/common_functions.glsl.js" should belong to a chunk
```

### Root cause

Regression from #10080. When a manual code-splitting group leaves a dynamic entry's facade chunk empty, `optimize_facade_entry_chunks` eliminates the facade and re-includes the entry's namespace object (`SymbolIncludeReason::SimulatedFacadeChunk`) so the namespace becomes an export of the merge-target chunk.

That include went through `drain_body_demand_stmts`, so it counted as *body demand* for the entry module and resurrected its deferred (gated) side-effect statements. In the reproduction, three-mesh-bvh's backwards-compatibility statement

```js
export const shaderIntersectFunction = ` ... ${BVHShaderGLSL.common_functions} ... `;
```

comes back to life, and its member expression canonicalizes through the `export *` chain straight to the glsl leaf module. The leaf therefore becomes included **after** `split_chunks` already assigned every included module to a chunk, so it belongs to no chunk and `compute_cross_chunk_links` panics.

### Fix

A simulated-facade include only materializes the eliminated facade's namespace object as a chunk export; it is not a semantic observation of the module, so it must not create body demand. `handle_include_symbol` now skips `drain_body_demand_stmts` for `SimulatedFacadeChunk` includes — mirroring the existing `WorkItem::Module` gate on `is_simulated_facade_chunk` a few lines below. This is safe because the namespace statement's symbol getters were already narrowed to link-time-used symbols before the replay, so the include cannot legitimately need anything that wasn't included at link time.

This also restores the invariant that chunk-layout choices don't change tree-shaking results: with the fix, the grouped build emits exactly what the ungrouped build emits (the dead `${common_functions}` statement no longer appears).

### Alternatives considered

Instead of (or in addition to) the semantic gate, the generate stage could detect modules newly included by the replay and assign them to the merge-target chunk. That converts the panic into working output, but it ships code the ungrouped build proves dead, and it papers over inclusion growing after chunk assignment instead of preventing it — so the semantic gate alone is preferred.

### Test

`crates/rolldown/tests/rolldown/issues/10337` mirrors the three-mesh-bvh shape: dynamic `import()` → an index module with `export * as` plus the deferred template-literal statement → an `export *` barrel → the glsl leaf, with `treeshake.moduleSideEffects: false` and a `codeSplitting` group. The entry only touches `res.MeshBVH`, so the leaf is reachable exclusively through the facade-elimination replay. The fixture panics on `main` and passes with this fix; it also executes the output under node and asserts the namespace member exists.