chore: run the assembled desktop E2E suite in CI

[Azganoth]

Summary

• Add a Desktop E2E job to ci.yml on the existing pull_request and push-to-main triggers, so assembled-boundary regressions — IPC, the real folder watcher, persisted state, the window close handshake — fail the run that causes them instead of waiting for a maintainer on Windows to remember the suite.
• Use the job shape Determine whether the desktop E2E suite should run in CI #184 measured: the pinned windows-2025-vs2026 label, timeout-minutes: 20, CARGO_PROFILE_DEV_DEBUG: 0, its own Swatinem/rust-cache entry for target/desktop-e2e, build and suite as separate steps, failure-only artifact upload at 7-day retention, and a teardown assertion for orphaned leafdown-e2e or msedgedriver processes and a listener on port 4445, sampled after ten seconds of grace for WebView2 host processes.
• Split test:e2e:desktop into build:e2e:desktop and a new test:e2e:desktop:run, so the workflow calls named scripts rather than hardcoding e2e/desktop/run.ts. The combined local command is unchanged.
• Save Rust caches only from main, in both this job and check-backend. Entries written during a pull request run are scoped to that pull request ref, so nothing but a re-run of the same branch can restore them while they consume the shared 10 GB budget and evict by LRU. Pull requests restore main's entry either way.
• Update CONTRIBUTING.md and docs/architecture.md: the suite stops being "Windows-local", it still requires Windows and still sits outside pnpm check, and the artifact sensitivity caveat now separates local runs from CI ones, whose paths are runner paths.

Related Issue

Closes #186

Verification

This job's real behavior is proven by this pull request's own run, which is the first execution of it.

Before pushing, .github/workflows/ci.yml was parsed to confirm the job structure and that the save-if expression survives YAML quoting in both cache steps, and the teardown script was extracted and parsed with the PowerShell AST parser. oxfmt --check and tsc -b pass across the repository.

Manually verified on Windows 11:

1. Ran both teardown probes against a clean machine; Get-Process -Name leafdown-e2e, msedgedriver and Get-NetTCPConnection -LocalPort 4445 -State Listen each return nothing without erroring under -ErrorAction SilentlyContinue.
2. Confirmed pnpm test:e2e:desktop and pnpm test:e2e:desktop:run both resolve after the script split.

Not verified: cold and warm timing against the 20-minute ceiling, the failure-evidence upload path, and the teardown assertion on a runner. The E2E cache key is new and save-if restricts saves to main, so every run before merge is a cold build — #184 measured that at 6m20s, inside the ceiling, but warm timing cannot be measured until the second main push after merge. The upload path can be exercised on demand with a workflow_dispatch run carrying LEAFDOWN_E2E_FORCE_FAILURE.

Notes

• actions/upload-artifact is pinned at v7, its current major, matching how the file pins actions/checkout and actions/setup-node.
• The save-if change reaches check-backend, which is outside Run the assembled-desktop E2E suite in CI #186's scope. It is applied to both Rust caches rather than one, because leaving the two jobs on different save policies is the kind of inconsistency that costs a future reader more than the change saves. It is isolated in its own commit if it should go through its own issue instead.
• Out of scope per Determine whether the desktop E2E suite should run in CI #184: WebdriverIO retries, trigger path filters, parallel workers, a platform matrix, and adding the suite to pnpm check. The standing response to a flaky failure is to re-run once to confirm, then fix the assertion to target durable state.