fix(e2e): stabilize test suite and fix blog revalidation regression - #420

Inflight prefetches buffer the entire response body (arrayBuffer) before
resolving. For routes with slow streaming loaders (e.g., parallel sidebar
with 5.5s loader), this blocks SPA navigation for the full loader
duration even though createFromFetch can consume streams incrementally.

When an inflight prefetch exists, skip it and do a fresh fetch instead.
The fresh fetch gives createFromFetch a streaming Response for
incremental consumption. The inflight prefetch completes in the
background and populates the cache for future navigations.

…h fetch"

This reverts commit 9699109.

ctx.url is stripped of _rsc* params by stripInternalParams() in the
middleware pipeline, so _rsc_partial, _rsc_segments, _rsc_action,
and _rsc_loader checks on ctx.url were always false in production.

This caused:
- No segment hash in cache keys (different partial navigations to
  the same URL with different client segments shared a cache slot)
- _rsc_action and _rsc_loader skip guards never firing (loader
  requests going through the document cache instead of being bypassed)
- mayNeedSSR() receiving stripped URL, falling back to Accept header
  heuristic instead of definitive _rsc_partial detection

Fix: derive _rsc* detection and segment hashing from ctx.request.url
(the original, unstripped URL). ctx.url is still used for pathname
matching and user-facing query params.

Also update test mock to simulate production behavior: ctx.url has
_rsc* params stripped, ctx.request.url has them intact.

Prefetch fetch was calling response.arrayBuffer() to fully buffer the
body before resolving the inflight promise. This blocked navigation
for the entire stream duration when reusing inflight prefetches for
routes with slow streaming loaders.

Remove the arrayBuffer() call — return the streaming Response directly.
Cache storage uses response.clone() so the original body stays
consumable by navigation's createFromFetch for incremental streaming.

…d actions

- Add parallelLoaderSources to ResolvedSegment for stable aggregate
  promise identity across renders
- Fix segment reconciler to preserve parallel loader fields during
  partial updates
- Add restoreParallelLoaderMarkers for cache-hit/revalidation paths
  where parallelLoading tag is lost after serialization
- Add e2e blog test covering action revalidation with parallel sidebar

…euse

Prefetch fetch now tees the response body instead of buffering with
arrayBuffer(). One branch goes to cache storage, the other resolves
the inflight promise for navigation consumption. This lets navigation
reuse inflight prefetches without waiting for the full body to download.

Restores the inflight prefetch reuse path in navigation-client that
was previously removed — it now works correctly because the inflight
promise resolves with a streaming Response (not a buffered one).

… combined/parallel-loader-streaming

…n' into combined/parallel-loader-streaming

…aming' into combined/parallel-loader-streaming

…ching, docs

Parallel entries are now keyed by slot name (Record<@slot, ParallelEntryData>)
instead of a flat array. This makes slot override semantics natural (record
key assignment = last-write-wins) and gives the resolution layer direct
slot→entry mapping.

- server/context.ts: ParallelEntries type (Partial<Record<@slot, ParallelEntryData>>),
  helper functions (getParallelEntries, getParallelSlotEntries, getParallelSlotCount)
- dsl-helpers.ts: parallel() creates one entry per slot, assigns to parent.parallel[slot]
- fresh.ts + revalidation.ts: iterate via getParallelSlotEntries(), dedup by entry ID
- segment-system.tsx: namespace-based restoreParallelLoaderMarkers and per-parallel
  loader isolation in renderSegments
- Fix stale-propagation test mock to include actual context exports

Docs:
- skills/parallel/SKILL.md: streaming behavior, slot override semantics
- execution-model.md: parallel streaming across all render paths
- feature-map.md: parallel slot streaming and override dedup
- feature-file-map.md: add segment-system.tsx and segment-reconciler.ts ownership
- semantic-change-checklist.md: parallel slots section with reconstruction invariant

…uning

- Split single `hmr` project into `hmr-loader` → `hmr-routes` → `hmr-intercept`
  chain to prevent concurrent file modifications (route-types-hmr was modifying
  route files while intercept-hmr expected them intact)
- Decouple HMR/webkit-smoke from dev/production dependencies — flaky dev tests
  no longer block 44+ HMR/webkit tests from running
- Set retries: 2 for both local and CI (was 0 locally)
- Set workers: 3 for both local and CI (was 6 locally) to reduce dev server
  contention that caused random test failures

…errides

- Set global expect timeout to 10s (was Playwright default 5s) to handle
  slower dev server responses under concurrent worker load
- Increase helper function default timeouts (waitForElement, clickAndWaitFor,
  waitForTextChange, waitForNumericChange) from 5s to 10s
- Remove explicit `{ timeout: 5000 }` from 43 positive assertions across
  8 flaky test files so they inherit the higher global default
- Increase pending-actions wait from 600ms to 1500ms for action settlement
- Increase handler-first cache write wait from 500ms to 1000ms
- Increase handle-breadcrumbs action timeout from 15s to 20s
- Increase timeout.test.ts waitForOnError default from 5s to 15s

Reduces flaky test count from ~17 to ~12 per run.

…d retries

- Increase waitForClientAgeToExceed polling timeout from 5s to 15s to
  handle slower timer increments under load
- Set retries: 2 for both local and CI (was 0 locally) to handle the
  flaky blog post query-param revalidation test

…ions

The blog probe action feature changed postRevalidation to use
`isBlogProbeAction ?? defaultShouldRevalidate` which falls through to
defaultShouldRevalidate (true) for non-action GET navigations with
search param changes. This caused the blog post to re-render on every
?tab=N navigation, resetting the client timer.

Fix: explicitly check for probe action OR slug change, preserving the
original behavior of skipping revalidation on search-only changes.

The hmr project was split into hmr-loader, hmr-routes, hmr-intercept
but the CI workflow still referenced the old 'hmr' project name,
causing CI to fail with "Project hmr not found".

…ze timing

- Remove clearLocalViteCache() and its call sites — isolated servers now
  use RANGO_E2E_VITE_CACHE_DIR env var instead of deleting shared cache
- Add default cacheDir fallbacks to vite configs (e2e-basic, e2e-timeout,
  test-app) so each app uses its own cache dir without the env var
- Add .vite* to .gitignore for isolated cache directories
- Small timing stabilization: waitForTimeout after hydration in cache
  and loader-behavior tests, expectNoReload guard in skipSSR test,
  increase prerender-intercept retry attempts from 2 to 5

After production eviction, Static() handler stubs lack a .handler
property, causing parallel slot resolution to skip (fresh path) or
crash with an invariant (revalidation path) before tryStaticSlot
could look up the pre-rendered component from __STATIC_MANIFEST.

Move tryStaticSlot before the handler undefined guard in fresh.ts,
and replace the invariant with a graceful fallback in revalidation.ts.

After `rm -rf node_modules/.vite`, the first HMR file change triggers
Vite dep re-optimization → full page reload, which destroys the
console event listener tests rely on. Add warmup tests that absorb
this initial reload so subsequent HMR tests get clean rsc:update events.

Isolated dev servers (use-cache, timeout tests) skip the shared
dev-warmup setup, so their first SSR request triggers Vite dep
optimization → module re-evaluation → in-memory cache loss. Add
warmupDevServer() that makes two consecutive SSR requests to absorb
the dep optimization cycle before tests run.

Also switch Unix kill() to process group kill (-pid) so Vite's child
processes (workerd, etc.) are cleaned up properly on teardown.