16 lines
9.6 KiB
Markdown
16 lines
9.6 KiB
Markdown
# Size change log
|
|
|
|
A running record of how each correctness/feature change affects the `WebComponent` base class bundle, measured as min + brotli via [size-limit](https://github.com/ai/size-limit). Baseline before the Phase 0 correctness work: **1.19 kB** (limit 1.2 kB).
|
|
|
|
| Change | Size (min+brotli) | Limit | Reason & benefit |
|
|
| ---------------------------------------------------------------- | ------------------ | ------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
|
|
| Move default-attribute reflection out of the constructor | ~1.19 kB (≈0) | 1.2 kB | **Custom Elements Spec compliance.** Constructors may not mutate attributes, so `document.createElement` on a class with `static props` threw `NotSupportedError` in real browsers. Defaults now reflect on first connect, and markup/SSR-provided attributes win. Pure code movement — no net size cost. |
|
|
| Safe prop cloning (functions/instances) + define-time validation | 1.31 kB (+0.12 kB) | 1.2 → 1.35 kB | **No more `DataCloneError`.** `structuredClone` of a function/class-instance default crashed construction. Defaults are now cloned per key (plain data deep-copied so instances don't share object/array state; non-cloneable values kept by reference), and a one-time warning names any default whose type can't reflect to an attribute. Limit raised to fit. |
|
|
| Derive prop types from defaults only; log type violations instead of throwing (+ `static strictProps` opt-in) | 1.33 kB (+0.02 kB) | 1.35 kB | **Loud, not fatal.** The proxy setter locked a prop's type on first write and threw on mismatch; inside `attributeChangedCallback` that `TypeError` vanished into `window.onerror` and silently skipped rendering. Types are now taken from `static props` defaults only (undeclared props stay untyped), and a declared-type violation is logged via `console.error` and skipped so `render()`/`onChanges()` still run. Teams wanting hard enforcement can set `static strictProps = true` to restore throwing. |
|
|
| Correct empty-string / removed attribute handling; drop native-setter writes | 1.33 kB (≈0) | 1.35 kB | **State no longer corrupts on common attribute changes.** `attr=""` coerced to boolean `true` (echoed as `attr="true"`); `removeAttribute` wrote `null` into the proxy and threw before `render()`/`onChanges()`; `deserialize('', 'number')` threw. `attributeChangedCallback` now keeps strings as-is (`''` stays `''`), resets removed attributes to the declared default, wraps `deserialize` so a malformed value falls back to the raw string instead of skipping render, and no longer writes vestigial `this[prop]` fields that clobbered native setters like `title`/`id`/`lang`. Logic swap — no net size cost. |
|
|
| `onChanges` gains a clear attribute-vs-property distinction (**breaking**) | 1.34 kB (+0.01 kB) | 1.35 kB | **Self-documenting change payload.** `onChanges` previously passed only `property`, holding the kebab-case _attribute_ name — confusing in a library where `props` are camelCase. `property` now holds the camelCase _prop_ key (matching `this.props` access) and the kebab attribute name moves to a new `attribute` field. Handlers can read `this.props[property]` directly without re-deriving the key. **Breaking:** code reading `changes.property` for the attribute name must switch to `changes.attribute` (see the migration note on the life-cycle-hooks docs page). |
|
|
| Buffer attribute-driven render/onChanges until after `onInit` | 1.35 kB (+0.01 kB) | 1.35 kB | **`onInit` is guaranteed to run first.** Per spec, authored attributes fire `attributeChangedCallback` (→ `render`/`onChanges`) _before_ `connectedCallback`, so components that assume `onInit` already ran break in real browsers (happy-dom hides this). A `#connected` flag now buffers the `render`/`onChanges` side effects until after `onInit`, while the prop value is still applied immediately so `props` is correct inside `onInit`. Pre-connect changes are reflected by the first render and not replayed through `onChanges`. |
|
|
| Reconcile `render()` in place instead of `replaceChildren` (WCB-REQ-06) | 1.81 kB (+0.43 kB) | 1.4 → 1.85 kB | **Re-renders no longer destroy DOM state.** The vnode path rebuilt the whole subtree and swapped it on every prop change, so focus, caret/selection, an uncommitted `<input>` value, `:hover` and running CSS transitions were wiped on the component's own rendered nodes — text fields lost focus mid-typing and a segmented control lost keyboard focus on select. Re-renders now reconcile the existing DOM against the new vnode tree in place (`src/utils/patch.mjs`: index-based, non-keyed — reuse same-tag elements, patch prop/attr adds, changes and removals by the same rule `createElement` applies, recurse children, trim extras). The first render still builds via `createElement`, the `JSON.stringify` no-op skip is unchanged, and `html`'s vnode shape and string templates are untouched. Reusing an element also means a `style` rule that is dropped or goes falsy (`{ fontStyle: condition && 'italic' }`) must now be cleared explicitly — a rebuild used to do that implicitly. The reconciler lives in its own zero-dep module, but `size-limit` measures each entry *with its dependencies*, so the core budget had to move regardless; no runtime dependency was added. |
|
|
| Reflect boolean props as bare attributes (WCB-REQ-07, **breaking**) | 1.93 kB (+0.12 kB) | 1.85 → 1.95 kB | **`toggleAttribute()` and `[attr]` CSS selectors work again.** Booleans reflected as the strings `flag="true"`/`flag="false"`, so the attribute always existed: `el.toggleAttribute('flag', true)` was a silent no-op (it only adds/removes, never rewrites a value — no `attributeChangedCallback`, no re-render), and presence selectors like `:host([flag])` matched the stamped `flag="false"` too, painting "on" styles onto elements whose prop was `false`. Boolean props now follow HTML in both directions: reflecting `true` sets the **bare** attribute, `false` **removes** it, and attribute removal means `false` rather than the declared default. Reads are strict — **any present value is `true`**, including the literal `flag="false"`, exactly like native `disabled="false"`. **Breaking:** `setAttribute(name, String(bool))` now always means `true`; migrate those call sites to `toggleAttribute(name, bool)`. Two dev warnings carry most of the cost and make the migration safe: one when a boolean attribute is written as a literal `"true"`/`"false"` string (the silent inversion), and one once-per-class when a boolean prop declares a `true` default (HTML has no true-default boolean attribute — absence must mean both "false" and "default"). `applyProp` also stops stamping `"false"` for boolean vnode props with no matching DOM property, which a nested `WebComponent` would otherwise read back as `true`. Limit raised to fit the warnings. |
|
|
| Overrideable `toAttribute` / `fromAttribute` converters (#56) | 1.96 kB (+0.03 kB) | 1.95 → 2 kB | **An escape hatch for the edge cases strict boolean semantics create.** The reflection and parsing rules were hardcoded, so a prop needing custom serialization (a `Date`, a delimited list, an enumerated `"true"`/`"false"` attribute) had no seam to hook into. Two overrideable methods now own both directions — `toAttribute(name, value)` returns the attribute value, where **`null` removes the attribute**, and `fromAttribute(name, value)` parses a present attribute back into a prop value. Both take the camelCase prop key, matching `static props` and `onChanges`'s `property`. The existing behavior became the default implementations (pure code movement — boolean presence/absence is now simply `toAttribute` returning `''` or `null`), so subclasses customize one prop and delegate the rest to `super`. Cheaper than a per-prop converter map, which would cost a lookup on every prop even for the vast majority of components that never need one. Limit raised for the two non-manglable public method names. |
|
|
| Don't reconcile a nested component's own light DOM | 2.01 kB (+0.05 kB) | 2 → 2.05 kB | **Nested components stop erasing each other.** The in-place reconciler recursed into every same-tag element it reused, including nested custom elements. A parent's vnode never describes what a nested light-DOM component rendered *for itself*, so the trailing-node trim in `patchChildren` deleted that content on every ancestor re-render — and since the child only re-renders when its own props change, it stayed blank. Any `WebComponent` that renders another `WebComponent` (without shadow DOM) lost the inner subtree the first time the outer one re-rendered, even for an unrelated prop. `patchNode` now treats a custom element that has no open shadow root as opaque: it still patches the element's props (that is how data flows down) but leaves the children to the child. Shadow-DOM components are exempt — their own content lives in the shadow root, invisible to the parent, so any *light* children are genuine parent-authored slot content that must still be reconciled. Limit raised 50 B to fit the one-line guard. |
|