…#10095)

### Description

Part 1 of the tree-shaking inclusion refactor stack.

`include_statements.rs` had grown to ~1200 lines interleaving the inclusion fixpoint driver, the three mutually recursive inclusion functions, dynamic-entry topological ordering, the body-demand gating model, and several standalone passes — with 10+ orthogonal policies (CJS bailouts, constant/enum inlining bypasses, JSON self-reference tracking, `eval` handling, HMR namespace registration, property-write retention) inlined directly into the recursion. Reviewing any change meant re-deriving which clause belonged to which concern; every bug fix added another inline branch.

Pure code motion, in three commits:

- **Split into focused modules**
  - `dynamic_entries.rs` — dynamic-entry topological ordering (tarjan), aliveness determination, per-iteration processing (~200 lines unrelated to inclusion mechanics)
  - `on_demand.rs` — the body-demand gating model (`side_effects_included_on_demand`, `compute_on_demand_side_effect_stmts`, `is_gated_side_effect_stmt`)
  - `passes.rs` — standalone passes around the fixpoint (CJS bailout export inclusion, runtime-helper collection/inclusion, `preserveModules` re-export interface preservation)
- **Extract named policies** — each inline block of `include_module` / `include_symbol` / `include_statement` becomes a named function with a doc contract (`sweep_side_effect_statements`, `include_side_effectful_dependencies`, `is_bypassed_inlined_constant`, `drain_body_demand_stmts`, `follow_cjs_namespace_alias`, `note_namespace_inclusion_reason`, `demand_esm_init_wrapper`, `note_json_self_reference`, `include_property_write_referencing_stmts`, `scan_import_records_for_cjs_bailout`, `is_inlined_enum_member_access`), so the three skeletons read as a sequence of policies. Also dedupes `include_declaring_statements`, which replaces five hand-rolled copies of the "a used binding keeps its declaring statements" loop.
- **wasm build fix** — the move dropped an incidental `itertools::Itertools` import that wasm targets need for `zip_eq` (`rolldown_utils::rayon` shims `par_iter()` to a plain `Iterator` there); caught by the browser/WASI CI jobs.

No behavior change: byte-identical output across the integration suite (zero snapshot diffs), clippy clean, `wasm32-wasip1-threads` checked.

Stack: #10095 → #10096 → #10097 → #10098

…st engine (#10096)

### Description

Part 2 of the tree-shaking inclusion refactor stack (stacked on #10095).

Replaces the `include_module` ↔ `include_symbol` ↔ `include_statement` mutual recursion with an explicit worklist engine. Edge producers push typed work items — `WorkItem::Module(idx)` (structural inclusion: sweep side-effect statements, evaluate side-effectful deps), `WorkItem::Symbol(ref, reason)` (a symbol became used: retain its declaration, wrapper, owner module), `WorkItem::Statement(idx, stmt)` (include one statement and follow its references) — onto a queue that the public `include_*` entry points drain to empty before returning. Handlers and internal helpers only push; only the public entry points drain, so drains never nest.

Why:

- no stack-depth limit on deep module graphs (the recursion could overflow)
- inclusion order becomes a visible queue discipline instead of an implicit call stack — the class of traversal-ordering bugs becomes inspectable and testable
- every inclusion flows through one loop, giving a single choke point for future reason/provenance tracking

Correctness argument: the final inclusion sets are a monotone closure over the pushed edges (all handler dedup is idempotent bit-setting; reason flags are ORed), so drain order affects only traversal order, never results. LIFO drain mirrors the depth-first shape of the recursion it replaces. Confirmed empirically: byte-identical output across the integration suite, zero snapshot diffs.

Public API is unchanged — `passes.rs`, `dynamic_entries.rs`, and `chunk_optimizer.rs` keep calling the same `include_*` functions, apart from the one new context field.

#### Review follow-up: queue-invariant hardening

The drain-before-return invariant made helper boundaries subtle — especially `include_declaring_statements`, which was enqueue-only yet callable from outside the engine (driver entry loop, dynamic entries), legitimately pre-seeding the queue across a public entry. Addressed:

- split it: `enqueue_declaring_statements` (private, engine-internal, used by handlers) vs `include_declaring_statements` (`pub(super)`, now drains like every other public entry) — so "queue empty between public entries" is a true invariant, not a mostly-true one
- `debug_assert!` at every public entry point that the queue is empty on entry and after draining; the full fixture suite runs with these live
- `WorkItem` and the `pending` field narrowed to `pub(in crate::stages)` (the minimum: `chunk_optimizer` constructs `IncludeContext` with a struct literal), documented as not-API

Stack: #10095 → #10096 → #10097 → #10098

…a stmt multimap (#10097)

### Description

Part 3 of the tree-shaking inclusion refactor stack (stacked on #10096).

#### Before

The body-demand machinery stored, for **every** body-demand key of a module, a full copy of that module's gated side-effect statement list (`FxHashMap<SymbolRef, Vec<(ModuleIdx, StmtInfoIdx)>>`), drained by map removal. That one mutable map entangled three different things: the **demand edge** (key → module), the **gate predicate** (which statements are gated), and the **progress state** (encoded as entry presence/absence). A module with N own exports carried N copies of the same statement list, and "has this module's body already joined?" was emergent from which keys happened to have been removed.

#### After

The rule being implemented — *a `sideEffects: false` module's gated statements join once its body is demanded* — is a boolean event on a **module**, and the state now says exactly that. Module inclusion is two explicit bits:

- `is_module_included_vec` — structural inclusion (wrapper/interop resolvable, side-effectful deps evaluated); unchanged
- `body_demand_swept` — the module's gated side-effect statements have joined because one of its own exports / its namespace object was actually used

and each of the three entangled concerns lives where it belongs:

- **Demand edge** — `body_demand_keys: FxHashMap<SymbolRef, ModuleIdx>` ("whose body does using this symbol demand?"), immutable and shared across both `IncludeContext` constructions (link stage and chunk-optimizer replay); no more `&mut` threading. Every key's canonical is owned by the module itself (own exports and namespace refs never link across modules). Before, `entry(key).or_default().extend(..)` would have silently merged two modules' statements if that invariant ever broke; now it's an explicit `debug_assert!` exercised by the whole suite.
- **Gate predicate** — the gated statements are enumerated at demand time from the immutable `stmt_infos`, via the same `is_gated_side_effect_stmt` used when building the key map. One source of truth; no O(keys × gated stmts) materialized copies.
- **Progress** — `body_demand_swept: FxHashSet<ModuleIdx>`, owned per run. `insert()` returning `false` short-circuits, so the sweep happens exactly once per **module**. Before, `remove(&key)` only guaranteed once per *key*: demanding a second export of the same module re-enqueued all its statements (harmless only because `include_statement` is idempotent — you had to know that to trust it). The generate-stage replay gets shared edges plus a fresh swept set: graph-scoped facts vs run-scoped progress.

#### Why `FxHashMap`/`FxHashSet` rather than `IndexBitSet`

`is_module_included_vec` is already an `IndexBitSet` — dense domain, touched for essentially every module on the hot path. The two body-demand structures sit on the other side of that line: they are sparse, usually **empty**. A module contributes keys only if it is user-declared side-effect-free, ESM, non-entry, without `eval`, **and** has a top-level side-effecting statement that reads module-level bindings — the rare #9806-family pattern.

- `body_demand_keys` can't be dense-indexed at all: `SymbolRef` is a two-dimensional key (`owner` × `SymbolId`), so a dense encoding would cost O(total symbols in the graph). The map *is* probed once per symbol in `include_symbol`, but a `get` on an empty `FxHashMap` is a few instructions — the common case stays effectively free.
- `body_demand_swept` is constructed fresh per `IncludeContext` (twice in the link stage, again in the chunk-optimizer replay). `IndexBitSet::new(modules.len())` would allocate and zero the full module range each time to record what is almost always nothing; `FxHashSet::default()` allocates nothing until the first insert, and it is only touched after a `body_demand_keys` hit. Its cardinality is bounded by the number of gated modules, so both structures are zero-cost together in the common case.
- Neither structure is ever iterated (membership test / probe only), so hash iteration order cannot leak into output.

#### Explored and rejected

Deriving `is_module_included` from the statement bits. The `SymbolIncludeReason::SimulatedFacadeChunk` path (chunk optimizer) deliberately includes statements **without** structurally including the module, so a derived bit is a behavior change, not a refactor.

No behavior change: byte-identical output across the integration suite, zero snapshot diffs.

Stack: #10095 → #10096 → #10097 → #10098

## [1.1.5] - 2026-07-08

### 🚀 Features

- detect top-level import-binding reads as execution-order sensitive (#10180) by @hyf0
- sourcemap_filenames: add a sourcemapFileNames option (#9271) by @V1OL3TF0X
- binding: record plugin hook result kind in tracing spans (#10154) by @IWANABETHATGUY
- linking: skip side-effect-free modules in per-entry reachability (#10111) by @IWANABETHATGUY
- improve error message for unresolved virtual imports (#10156) by @sapphi-red
- add descriptive metadata to plugin API (#10106) by @sapphi-red
- add `--configLoader=native` option (#10118) by @sapphi-red

### 🐛 Bug Fixes

- improve invalid annotation warnings (#10185) by @hyf0
- keep deduplicated asset filenames stable once they can be observed (#10191) by @shulaoda
- sourcemap_filenames: use public option name in pattern errors (#10188) by @IWANABETHATGUY
- sourcemap_filenames: hash prepared sourcemap content (#10178) by @hyf0
- tree-shake unused circular declarators exported via export list (#10166) by @IWANABETHATGUY
- dev: don't panic when an HMR rebuild hits an unresolved import (#10162) by @shulaoda
- propagate errors from output.globals function (#9880) by @shulaoda
- dev: revert cache mutations when a partial scan fails (#10110) by @shulaoda
- dev: update importer relationships of cached modules in incremental build (#10107) by @shulaoda
- hmr: fall back to full reload when a changed module is not registered as executed (#10132) by @shulaoda
- chunk-optimizer: follow entry facade edges in runtime placement cycle check (#10101) by @hyf0
- dev: ignore watcher events after close (#10113) by @hyf0
- emit async wrapper for TLA modules under onDemandWrapping (#10086) by @IWANABETHATGUY
- gate sideEffects:false modules' side effects on body demand (#10080) by @IWANABETHATGUY
- rolldown_plugin_vite_resolve: return empty object for `browser: false` mapped modules (#10082) by @sapphi-red
- reset the word-boundary state on newline in Hires::Boundary sourcemaps (#10025) by @shulaoda
- trim an emptied chunk's outro/intro instead of skipping past it (#10029) by @shulaoda
- test each edited chunk's own start against indent exclude ranges (#10026) by @shulaoda
- preserve sourcemap mappings for indented lines when a CJS module shares the chunk (#10074) by @hyf0

### 🚜 Refactor

- separate tree-shaking side effects from execution order sensitivity (#10168) by @hyf0
- type construct_vite_preload_call to take an ObjectPattern (#10135) by @shulaoda
- treeshake: single-source the own-export classification shared with the lazy-barrel loader (#10098) by @IWANABETHATGUY
- dev: reuse Vite's bundledDev server (#10081) by @h-a-n-a
- clippy: ban std HashMap/HashSet in favour of FxHashMap/FxHashSet (#10108) by @Boshen
- treeshake: make body demand a second module bit instead of a stmt multimap (#10097) by @IWANABETHATGUY
- seal used_symbol_refs by construction after its last writer (#10091) by @hyf0
- treeshake: replace inclusion mutual recursion with a worklist engine (#10096) by @IWANABETHATGUY
- treeshake: split include_statements.rs into focused modules (#10095) by @IWANABETHATGUY
- drop redundant is_user_defined filter on partitioned entries (#10050) by @shulaoda
- project the retained export interface out of used_symbol_refs (#10089) by @hyf0
- track used external symbols separately from used_symbol_refs (#10088) by @hyf0
- make module namespace inclusion an explicit linking metadata field (#10087) by @hyf0
- rename statement evaluation metadata (#10078) by @hyf0

### 📚 Documentation

- virtual modules user-facing id convention (#10155) by @sapphi-red
- cli: clarify disabling boolean/object flags like codeSplitting (#10153) by @IWANABETHATGUY
- chore: remove Vite+ alpha banner (#10105) by @mdong1909
- write down the used_symbol_refs contract (#10090) by @hyf0
- dev/lazy: update design and implementation (#10079) by @h-a-n-a

### ⚡ Performance

- ast_scanner: stop order-sensitivity checks once a module is flagged (#10190) by @IWANABETHATGUY
- return impl ExactSizeIterator from slice-backed accessors (#10133) by @Boshen
- binding: box dev and watcher napi futures (#10103) by @Boshen

### 🧪 Testing

- move string_wizard replace unit tests to the JS magic-string suite (#10176) by @IWANABETHATGUY
- dev: assert incremental scan state matches a fresh full build after each HMR step (#10115) by @shulaoda
- dev: restore runtime assertions of delete_file_not_used_anymore (#10112) by @shulaoda
- dev: fix flaky dev server tests in CI (#10152) by @h-a-n-a
- add regression test for #10099 (lazyBarrel drops default-import binding but keeps its property reads) (#10109) by @IWANABETHATGUY

### ⚙️ Miscellaneous Tasks

- deploy website to Void via GitHub OIDC (#10192) by @Boshen
- deps: update oxc to 0.139.0 (#10161) by @shulaoda
- deps: update test262 submodule for tests (#10160) by @rolldown-guard[bot]
- rolldown_plugin_utils: remove dead asset-url and css scaffolding (#10131) by @shulaoda
- deps: revert vite-plus to v0.2.1 (#10148) by @shulaoda
- deps: update github actions (#10141) by @renovate[bot]
- deps: update dependency rust to v1.96.1 (#10145) by @renovate[bot]
- deps: update npm packages (#10142) by @renovate[bot]
- deps: update rust crates (#10143) by @renovate[bot]
- deps: update napi to v3.10.3 (#10121) by @renovate[bot]
- rolldown_utils: remove unused time module (#10138) by @shulaoda
- remove dead CopyModulePlugin::is_active method (#10129) by @shulaoda
- remove dead LazyCompilationContext::is_lazy_module method (#10128) by @shulaoda
- remove dead BuildDiagnostic::downcast_ref method (#10127) by @shulaoda
- deps: update dependency vite-plus to v0.2.2 (#10084) by @renovate[bot]
- deps: update rust crate oxc_sourcemap to v8.1.0 (#10122) by @renovate[bot]
- deps: update crate-ci/typos action to v1.48.0 (#10124) by @renovate[bot]
- enable more clippy restriction lints (#10114) by @Boshen
- deps: update rust dependencies (#10100) by @Boshen
- deps: update oxc resolver to v11.23.0 (#10083) by @renovate[bot]

### ◀️ Revert

- Revert "chore(deps): revert vite-plus to v0.2.1" (#10157) by @h-a-n-a
- "fix(hmr): fall back to full reload when a changed module is not registered as executed (#10132)" (#10151) by @shulaoda

### ❤️ New Contributors

* @V1OL3TF0X made their first contribution in [#9271](#9271)

Co-authored-by: shulaoda <[email protected]>