)

### What this PR solves

Any wrap trigger — `strictExecutionOrder`, `require()` of ESM, or an inlined dynamic import under `codeSplitting: false` — force-includes a `sideEffects: false` ESM module through its wrapper ref, retaining binding-reading side-effect statements that plain (unwrapped) tree-shaking provably drops. With `experimental.lazyBarrel`, those retained statements reference import records the loader correctly never loaded (it defers records by requested exports), so the emitted chunk contains free identifiers and throws `ReferenceError` at load. The tree-shaking inconsistency between `strictExecutionOrder: true` and `false` is the shared root cause of the family:

fixes #9691, fixes #9806, fixes #9961, fixes #9964, fixes #10013, fixes #10048

### How

A binding-reading side-effect statement of a user-declared side-effect-free ESM module is now included **iff the module's body is demanded**: one of its own (non-re-export) exports or its namespace object becomes used. This mirrors the lazy-barrel loader's `local`/`All` classification exactly — the loader loads *every* plain import record precisely in those cases — so a kept statement can never reference an unloaded record, and a deferred record is only ever referenced by dropped statements. The loader itself needs no changes.

Behavior deliberately kept as-is (pinned by existing tests):

- statements with no module-symbol references (a bare `console.log()`) and import/re-export statements (they drive wrapper `init_*` chains and side-effect imports) keep the unconditional sweep — `package_json_side_effects_false_keep_*` esbuild fixtures;
- bodies of modules whose own export is demanded (#7597's top-level asserts) or whose namespace is observed (`import * as ns`, `require(esm)`) are retained unchanged;
- user-defined entries, `eval` modules, and CommonJS modules are exempt.

Dynamic entries are **not** exempt: a live dynamic entry's body joins through namespace/own-export demand, while a dead pure dynamic import must not resurrect a body its removal already models as an empty namespace (covered by the new `tree_shaking/dead_dynamic_entry_side_effect_free` fixture).

### Alternatives explored

- Loader-side fix (force-load records referenced by retained statements): rejected — the deferral strategy is correct; the retention is the bug.
- Blanket statement-level stripping for `sideEffects: false` modules: rejected — breaks the esbuild-parity `keep_*` fixtures (namespace/require demand must keep the body).
- Statement-granular demand edges (keep a statement once any symbol it references is used): rejected — still dropped #7597's asserts.

### Notes for reviewers

- Inclusion changes only for `UserDefined(false)` ESM modules reached exclusively through re-exports or wrapper refs; every pre-existing snapshot is byte-identical (full integration suite: 1785 passed, 0 failed).
- Each `issues/*` fixture executes its output with node; `issues/10013` and the dynamic-entry fixture additionally carry a local `node_modules/dep` that throws on evaluation, so re-emitting the dropped external import fails execution, not just the snapshot.
- Residual and deferred (benign): wrapped modules whose body was pruned still emit empty `init_x` shells; a follow-up no-op-init cleanup can drop them.