fix: rc.10 follow-ups — live server-function address in dev, event-stream preview passthrough, attribution/perf-tracks pre-bundling

[ryansolid]

Four follow-ups found by the rc.9 → rc.10 audit of the plugin against Solid 2.0.0-rc.10. No plugin option added, removed or renamed.

1. live/ server-function address in the dev middleware

What. The dev middleware's id parser strips a leading live/ segment the way it strips data/, mirroring the runtime's parseServerFunctionAddress (one kind segment at most; an id is exactly one segment, so live/, live/a/b and data/live/<id> stay misses).

Why. rc.10 added a third address shape, <endpoint>/live/<id>, for a live loop's connection (answered as text/event-stream). The parser only knew data/, so live/<id> kept its slash → functionId = null → the on-demand SSR import was skipped → a live function whose module nothing had imported server-side yet came back 404 Unknown server function (reproduced before the fix; 200 event-stream after).

Where. src/server-functions/index.ts:649 (the strip), comments at src/server-functions/index.ts:611, :639 and src/ssr/index.ts:1324 now enumerate all three shapes.

2. Preview compression vs event streams

What. vite preview's handler middleware sets content-encoding: identity for text/event-stream responses and for everything under the server-function mount, in addition to text/html.

Why. @polka/compression sizes a response on its first write and, at ≥1KiB, pipes the rest through zlib, which holds small writes until its buffer fills — a live event stream's events would sit undelivered while the connection stays open; the streaming codec's chunks likewise arrive in bursts. Rule chosen and documented in the code: content type for text/html and text/event-stream (SSE from an app middleware route is a stream too), plus the server-function mount, because the runtime's other streamed shape — the codec stream — is text/plain, too generic to key on, and keying on the mount keeps a future answer shape from silently regressing to buffered. The mount reaches the start plugin through startServe's existing internal wiring (serverFunctionsEndpoint; startServe is not part of the package's exports), normalized by a shared normalizeServerFunctionsEndpoint (same logic the server-functions plugin already used inline).

Where. src/ssr/index.ts:1824–1885 (mount + rule), src/ssr/index.ts:630 (internal field), src/index.ts:2143 (threading), src/server-functions/index.ts:182 (normalizer, used at :382).

3. optimizeDeps.include: solid-js/attribution, @solidjs/web/performance-tracks

What. Both subpaths join optimizeDeps.include under serve — the same gate the refresh runtime and the frames/server-functions entries already use; not gated on the dev/observe posture (the surrounding entries aren't, and the subpaths exist in every posture).

Why. A consumer the scanner never crawls (a linked package, a solid-condition package vitefu excluded) importing one mid-session triggers late discovery → re-optimize → a second @solidjs/signals core beside the one solid-js was bundled with. Verified both subpaths exist in packages/solid/package.json / packages/web/package.json on origin/next and at tag solid-js@2.0.0-rc.10 (./attribution since rc.8, ./performance-tracks since rc.10; peer range is ^2.0.0-rc.10).

Where. src/index.ts:1382–1394.

4. Stale strings

• README.md:124 and src/index.ts:265 no longer say the dev option "injects solid-js/dev" (no such entry); it resolves the development export condition.
• examples/start-ssr/vite.config.ts:35, examples/start-ssr/test/run.mjs:4420, :4453: componentNames → sourceNames.

Tests

• examples/start-ssr/test/run.mjs:1001–1051 (dev mode): cold dispatch at /_server/live/<id> for incrementCounter in src/frames/data.tsx — a module App.tsx never imports and the bare cold dispatch didn't load — expects 200 text/event-stream with a data: event; plus the three grammar misses. Fails on origin/next (404), passes here.
• examples/start-ssr/test/run.mjs:3376–3468 (preview mode): greet at the live address (201 event-stream, first chunk >1KiB) and getServerMessage at the data address (text/plain) come back without gzip/br, while the hashed client asset still compresses. With the old text/html-only rule both come back gzip (verified).
• examples/start-ssr/test/components-warning.mjs:120–140: the two entries present under serve, absent under build.

Results (all local, macOS):

suite	result
pnpm build (rollup + tsc --emitDeclarationOnly)	ok
pnpm check	100/100
root pnpm test (vite-8 Cypress)	ok
examples/start-ssr pnpm test (run.mjs all modes + http-bridge + components-warning)	607/607, 10/10, 11/11
examples/start-client	65/65
examples/start-env	47/47
examples/ssr	12/12 + 8/8 boundary
examples/css-matrix	87/87 + 19/19 bridge
examples/vite-8 vitest (browser)	1/1

Prettier: the touched files carry pre-existing drift on lines this PR doesn't touch; every line this PR adds or changes is prettier-clean.

Public API changes: none

startServe's internal parameter gained serverFunctionsEndpoint, and normalizeServerFunctionsEndpoint is a new module-level export of src/server-functions/index.ts; neither is reachable from the package's export map (. re-exports serverFunctions and types only).

[changeset-bot]

🦋 Changeset detected

Latest commit: e40e2c4

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
@solidjs/vite-plugin	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[pkg-pr-new]

Open in StackBlitz

npm i https://pkg.pr.new/@solidjs/vite-plugin@374

commit: e40e2c4