[adhikjoshi]

Problem

Illuminate\Foundation\Vite is a container singleton on the shared base application, so every coroutine in a worker shared one $preloadedAssets list. Blade's @vite appends to it, AddLinkHeadersForPreloadedAssets builds the Link header from it, and FlushVite (RequestReceived) empties it whenever another request starts. Under concurrency a response's Link header carried another page's preload links, or none.

v0.8.34 (#22) made every Link value reach the client, which exposed this. Production impact (modelslab.com, 2026-10-15 19:45 UTC): 23 of 60 parallel fresh renders had a wrong preload header (7 empty, 12 with other pages' assets, 4 mixed). The app's full-page cache stores that list on a MISS and replays it for 60–300 s.

Fix

RequestScope resolves Illuminate\Foundation\Vite per coroutine: a clone of the boot-time instance (build directory, nonce, resolvers, prefetch settings are kept) with an empty preload list. @vite (app()), the Vite facade (caching is off in coroutine mode) and FlushVite all go through the scope.

The list is reset directly instead of with Vite::flush(), which only exists from Laravel 12.10 — calling it would throw on Laravel 10/11/12.0–12.9 and, because FlushVite resolves Vite on every request, turn every request into a 500.

Verification

• Unit (tests/Unit/RequestScopeViteIsolationTest.php): the clone keeps boot config and starts empty; and the production interleaving — request A renders, request B starts (FlushVite) and renders, A builds its header — read through the Vite facade and app() with the worker's global wiring. Both tests fail with the fix removed.
• Matrix (Swoole 6.3.0RC1, PHP 8.3.35): full suite green on Laravel 11.57 (no Vite::flush), 12.39 and 13.36.
• Real app under Swoole (modelslab-frontend-v2 production image at a0b8a39e0, Octane with the prod flags — 8 workers, 2 task workers, --max-requests=500 — against throwaway MySQL/Redis/Mongo; this branch bind-mounted over vendor/.../src):
  • 16 pages rendered fresh, 16 in parallel: v0.8.34 84/160 wrong headers; this branch 0/480 wrong (every header equals the page's own @vite assets).
  • Cost: clone + reset 0.063 µs/request; simulated request lifecycle 4.8–5.1 µs vs 5.5 µs before (noise).
  • Memory: 100k simulated requests → 0 bytes growth. Real app with recycling off (mixed load, 32–36k requests): v0.8.34 +9.7 KB/request, this branch +10.0 KB/request — same pre-existing growth, nothing added.
  • wrk (16 connections): within run-to-run noise (±20% on this machine); with order reversed the branch was faster.

Known limits (unchanged by this PR)

• A binding cached on the base container (e.g. a scoped() CSP-nonce generator) is still shared across concurrent requests; a nonce set on one request's Vite clone is not seen by another. Before this PR the nonce lived on the shared Vite and raced instead.
• An app that aliases Illuminate\Foundation\Vite to its own class, or injects Vite into a boot-time singleton, still gets the shared instance.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}