fix(checker): JSX children become part of the component contract

[SupulHeshan]

Two defects, one root cause: a component receives only what it declares. Children handed to a component that never declares children were dropped silently, a blank render behind a clean jac check. Fixes #8260, follow-up to #8134.

The mechanism

The client codegen destructures a component's declared parameter names out of props with no rest element, so anything undeclared has nowhere to land.

flowchart LR
  CS["call site passes<br/>title + nested children"] --> P["props = { title, children }"]
  P --> D["emitted:<br/>const { title } = props"]
  D --> OK["title bound"]
  D -.->|"no rest element"| X["children dropped<br/>no runtime signal"]

Loading

The codegen is deliberately left as is. This PR makes the contract visible instead of passing children through, so a component still receives exactly what it declares.

The deeper gate

The checker never looked at the call sites where this shows up. _validate_jsx_props returned early when the call site passed no attribute, and _validate_component_jsx_props returned early when the component declared no parameter. So <Card/> against def Card(title: str) checked clean, <Zero bogus="x"/> against a zero-parameter component checked clean, and <Card><Leaf/></Card> never reached validation at all. Both early exits are removed; every call site is validated like any other. This can surface E1102 and E1101 on code that previously passed.

The contract

flowchart TD
  A{"return type names JsxElement,<br/>JsxPage or JsxLayout?"}
  A -->|no| Z0["not a component, nothing checked"]
  A -->|yes| B{"parameter shape"}
  B -->|"'props' plus other params"| E9["E1109 error<br/>on the declaration"]
  B -->|"'props' alone"| Z1["fine, children arrive<br/>as props.children"]
  B -->|"declares 'children'"| Z2["fine"]
  B -->|"no 'children'"| C{"who passes the children?"}
  C -->|"a Jac call site"| W["W1053 warning<br/>at the call site"]
  C -->|"the generated pages/ entry"| E8["client build refuses<br/>the entry"]

Loading

Where	Rejects	Why there
W1053 (checker)	children passed to a component declaring none	the call site is visible, so the author can see and fix it
E1109 (checker)	a props bundle beside another parameter	emits either invalid JavaScript or a signature the renderer never calls that way
ClientBundleError (bundler)	a pages/ app that would discard the whole route tree	the call site is generated JS the checker never sees; the bundler has the scanner result, the app root and the compiled manifest's parameter names, so it refuses before generating the entry, and a warm build stays a cache hit

Nested content counts toward a children parameter only when something is rendered: whitespace and {#* comments *#} do not count, for W1053 and for E1102 alike.

Component-hood comes from one predicate shared with the codegen, return_type_is_jsx in symbol_utils, so JsxPage, JsxLayout and unions are covered exactly as JsxElement is.

E1109 concretely:

// def Mixed(props: any, tone: str)  ->
function Mixed(props) { const {props, tone = "flat"} = props; }  // SyntaxError

// def MixedKw(props: any, *, tone: str)  ->
function MixedKw(props, tone = "flat") { }  // renderer passes one object; tone is always "flat"

Solid: element= parent routes

The mirror image of the React defect, and the same blank page with no error anywhere.

flowchart TD
  subgraph B["before"]
    B1["Route with element={Guard}"] --> B2["wrapped in a<br/>props-ignoring lambda"]
    B2 --> B3["matched child never reaches the guard"]
    B3 --> B4["parent-route guard renders nothing"]
  end
  subgraph A["after"]
    A1["Route with element={Guard}"] --> A2["wrapper provides the matched child<br/>through the outlet context"]
    A2 --> A3["AuthGuard falls back to Outlet<br/>only when given no children"]
    A3 --> A4["parent-route and layout forms both render"]
  end

Loading

Also in this PR

• W1052 (untyped props: any bag) now uses the shared predicate, so it also fires on components returning JsxPage, JsxLayout or a union; before it fired on JsxElement only.
• _jac.dict.eq in the JS runtime guards null and non-object operands. The ES backend lowers x != None loosely but not x != undefined, so a dict compared to undefined reached Object.keys(undefined) and threw. Found by the Solid jsdom test.

Tests

Cases live in the tests that already own the behaviour; the diagnostics extend existing checks rather than add new areas.

Test	Asserts
jsx_step3_jsx_prop_check (extended)	E1102 on <Plain/> and on a required children with nothing or only a comment nested; E1101 on a zero-parameter component; W1053 on two components declaring no children; nested content satisfies children
jsx_props_bag_untyped_and_typed (extended)	one E1109 on props beside another parameter
union JsxElement return type still triggers props-destructure lowering (extended)	JsxPage and JsxLayout components lower through the props protocol like unions do
pages entry that ignores children drops every route (route scanner)	the manifest predicate across no parameters, other parameters, children, lone props, and no client app
a pages entry whose app cannot receive the route tree is refused (client warm start)	the real client compiler raises on a routed project whose app takes no children
solid file-based routing mounts, navigates, and guards in jsdom (extended)	four manual-routing scenarios through a second entry: /nested logged in and out, /layout-guarded, /bare

Docs: jac-cl-components.md gains the children contract and both codes, jac-cl-routing.md's "silently drops every route" footgun becomes the build refusal it now is, jac-cl-auth.md gains the Solid direction of the same bug, and the components tutorial's "children must have a default" warning becomes a tip that says why.

Not in scope

React-native AuthGuard render coverage (issue task 4, second bullet). react_native_multiscreen's test still asserts only returncode == 0 plus an APK directory; a real render assertion needs a react-native renderer in CI.

[christianwilkins]

CI audit from current main:

• Contribution Checks is patch-specific: this PR needs release_notes/unreleased/jaclang/8314.<type>.md (or the skip label).
• The substantive failures in test-compiler, test-runtime, test-client, test-packages-and-docs, native macOS, and the scale lanes all reduce to the same stale builtin surface (testskip / testraises missing, plus JCIR_BUILTIN_NAMES drift). The run started before feat(compiler): sixteen sealed roots, the pass rim, and the type system's first census (#8288) #8316 merged; this branch is now 21 main commits behind.

A rebase/update onto current main plus the release-note fragment should clear the observed blockers before rerunning CI. I also checked the patch against #8260's checklist: it covers the component children contract, pages/ app contract, Solid element= child forwarding, and the requested integration/CI coverage, so a duplicate PR would not add value.

[christianwilkins]

Follow-up on the current head 0c8d4b3c4af7: the release-note and stale-main CI blockers from my earlier audit are resolved. All 26 active checks pass, with two expected skips. GitHub now reports the branch as DIRTY against main, so resolving the current merge conflicts is the remaining merge blocker; there is no failing check at this head.

[kugesan1105]

---

[kugesan1105]

I pushed six commits onto this branch. Here is what each one changes and why, in plain terms.

1. Merge main into the branch

2. Drop the leftover if expr.children line in the checker
Main had added a small partial fix that counts any child as children. The rebase kept it beside this PR's stricter _has_rendered_jsx_children, which ignores comments and whitespace. The two disagreed: <Card title="ok">{#* note *#}</Card> against a required children parameter got no E1102, while W1053 correctly said nothing was passed. Deleted the old line; one fixture line now locks the case.

3. Fold the new tests into the tests that already own the behaviour
The new diagnostics extend behaviour that main already tests, so the cases moved there instead of living in parallel fixtures:

• E1101/E1102 at attribute-less and zero-parameter call sites, and W1053, joined the step-3 JSX prop-check fixture. Required-prop checking had no test before, which is how the attribute-less gate went unnoticed.
• E1109 joined the props-bag fixture that already declares both bundle styles.
• JsxPage/JsxLayout lowering joined the union-return codegen fixture. The predicate is shared, so one spelling each is enough.
• The Solid element= guard scenarios run through the existing Solid routing jsdom test with a second entry module, instead of a third copy of the jsdom boot and install loop.
• Dropped: the three parallel checker fixtures, the assertion about three-node JSX parsing (the parser has been single-node since refactor(compiler)!: type checking joins the critical path, and the optional gate burns down (#8398) #8399, so it could never fail), the /wrapped scenario that passed before the fix, and the React pages/ jsdom test, which exercised code this PR does not change.
  Net for jac/tests: +142 / -739.

4. Move the shared JSX predicate into symbol_utils
jaclang/compiler/jsx_component.jac was a two-function module at the compiler package root. symbol_utils.jac is where the compiler already keeps shared unitree helpers, and both the ES backend and the ownership checker import it. Moved the functions there and deleted the file.

5. Inline the runtime-helper and return-type literals
is_jsx_runtime_helper was a one-line wrapper around name.startswith("__jac") used at two sites, and JSX_RETURN_TYPE_NAMES a glob for a three-name tuple used at one. Both are written where they are read now. Only return_type_is_jsx, the union-aware predicate the checker and the emitter share, stays in symbol_utils.

6. Move the pages-entry check from the checker to the bundler
E1108 had the type checker answer "is this module the entry of a routed pages/ project" by itself: find the nearest jac.toml, list pages/*.jac, parse each one. That needed the new jaclang/project/pages_routing.jac module (the checker cannot import the client scanner) and it used a narrower rule than the real scanner: bare JsxPage names only, project root only, jac.toml required. It failed open on -> JsxPage | None pages, toml-less projects, and the workspace apps that landed in #8823.
The bundler already has the scanner result and the app root when it generates the entry that renders app with the routes as children, and the compiled manifest already records every exported function's parameter names. So JacClientCompiler._scan_and_compile_pages now reads manifest.params["app"] and raises a ClientBundleError with the same explanation when there is no children parameter and it is not a lone props. Reading the manifest matters: a first version compiled the entry module and broke the warm-build test, because a warm build must compile nothing.
Removed with it: E1108, pages_routing.jac, the checker-side function, its fixture projects and tests. The route scanner is back to main's version. The warm-start test's fixture app gained a children parameter, because it really was dropping the routes it declared.

Net for the PR against main: 31 files, +469 / -93, down from +1189 / -86. Every diagnostic and the Solid fix is still proven by a running test (checker JSX 9, codegen 5, route scanner 13, client warm start 2 including the end-to-end refusal, mobile target 54, Solid routing jsdom with 7 scenarios).

[kugesan1105]

Re-validated the final head live against main 0.37.7 (checker probe on 20 call-site shapes vs the same probe on main, bundler probe on ten project shapes including toml-less, JsxPage | None pages, Solid, and a [apps.web] workspace app, warm rebuild from the cached manifest). Everything holds, and the targeted suites pass locally.

One residual, filed as #9015 so it is not lost with the merge: the hot-reload path recompiles the edited module without running the new entry check, so editing app(children) to app() under jac start hot-swaps a route-dropping app silently until the next full build. Small follow-up, not a blocker. Merging.