Add isolated Redis e2e suite for registerInitialCache and handler-level correctness

[AyronK]

Summary

Adds a self-contained Playwright suite (examples/redis-minimal/e2e/isolated/) that manages its own ephemeral Redis container(s) and next start process(es), independent of the existing shared e2e suite's server/Redis. It proves behavior the shared suite can't reach because it only observes an already-running, already-mutated server:

• registerInitialCache actually populates Redis on boot — asserted directly against Redis before any HTTP request is made, for both App Router and Pages Router (fallback: false) routes, plus fetch-cache-derived tags.
• setOnlyIfNotExists behaves correctly in both directions — a pre-existing key survives when the flag is true and is overwritten when left at its default false, using the identical seeded input in both cases (a real A/B control, not just a single-condition observation).
• sharedTagsKey/sharedTagsTtlKey population rules and real Redis TTL, tested via direct construction of the redis-strings handler (the same public @fortedigital/nextjs-cache-handler/redis-strings subpath export the example app's own cache-handler.mjs uses) against a real Redis instance — no next start needed, so it can cover a full matrix of revalidate values and the lifespan: null permanent-key case that's unreachable via any real route at runtime.
• revalidateTag mechanics — selective deletion (untouched entries proven byte-for-byte unchanged), safe no-op on no match, and the non-obvious distinction between an implicit tag (_N_T_...) that matches an existing entry (eagerly deleted, same as any tag) versus one that doesn't (only bookkeeping is written — the genuinely lazy path) — plus the TTL-expiry sweep that runs as a side effect of every revalidateTag call.
• Full pipeline bridge — one test drives revalidateTag through the real /api/revalidate HTTP route and asserts the Redis deletion directly, rather than inferring it from a timestamp change.

Also:

• Splits the CI redis-minimal-e2e job into two parallel jobs (main suite + isolated suite), each fully self-contained.
• Adds JUnit-based test reporting (dorny/test-reporter) to both integration-tests.yml and node.js.yml, surfaced as GitHub Check runs.
• Fixes two real flakes surfaced along the way: a per-attempt Redis connect timeout (a hung TCP handshake right after docker run -d -p could swallow an entire retry budget on Linux CI), and a port-collision race between sequential spec files (fixed by allocating a genuinely free OS-assigned port per container/server instead of sharing hardcoded ports).
• Documents the two testing styles (boot next start vs. direct handler construction) in the new-e2e-test skill for future additions.

Test plan

• tsc --noEmit clean
• Full isolated suite passes locally (25/25), run multiple times to check for the port-collision flake
• Every new/changed assertion mutation-tested (flipped to a wrong expected value, confirmed a real failure, reverted) — including the subtlest ones (implicit-tag bookkeeping-only path, TTL cross-checks)
• Root-level pnpm test:e2e:isolated verified to correctly trigger the Turborepo build chain from a clean .next/dist (this is what CI actually invokes)
• No leftover Docker containers after any run
• CI green on this PR (parallel jobs + JUnit reporting)