test: prune low-value tests and parallelize the integration pool - #27

The runtime-playground HMR e2e test still stamped a fixed 30s Playwright
budget tuned on many-core machines onto waits that sit behind rsbuild
compiles and Chrome sharing a two-core runner - the same shape that
tripped the capture script on the Node 22.19 Verify job. Both surfaces
now read the shared timeScale helper instead of an inline CI multiplier,
so the two-core rationale lives in one documented place. Scaling costs
nothing on green runs since every wait returns on success.

… and budget the retention test for CI

…and rsc example

Removes tests that re-assert coverage pinned elsewhere or assert file
contents instead of behavior: the artifacts-real e2e file (its table
coverage lives in overview), the 390px Runtime controls e2e, the safe
launch configuration overview test (redaction lives in the epoch MCP
session test), the handoff close-retry overview test (pinned by
runtime-mcp-handoff.test.ts), config/source string-matching tests, and
the rsc example's tsconfig/doc/manifest duplicates. Also trims the
Inspector-tab detours from the Runtime sibling e2e ahead of the
inspector removal.

…suites

The Workbench is a desktop-only product validated at 1440x900, so the
390px viewport resizes and horizontal-overflow checks appended to real
host e2e tests assert a layout the product does not ship. Removes only
those mobile assertion sites, keeps every host test, runs the MCP App
preview browser test at the desktop viewport, and deletes the mobile
capture path (PNG, mobileLayout evidence, --mobile flag) from the
runtime playground capture script plus the README/topology command
strings that pin its exact invocation.

Each eviction test prepared its own dev session and drove fifty real
invocations before exercising one hook, repeating the slowest setup in
the suite four times (and flaking under load). One session now fills
the fifty-artifact window once and drives the happy eviction, the
held-reader reservation, the failed run-directory removal, and the
failed artifact release in eviction order, preserving every distinct
assertion including the readRunFlight path-traversal check and the
close-retry accounting. Neighbouring retain-until-close, worker-bound,
and containment tests are untouched.

…roduction build test

The second multi-environment rebuild test repeated the full example
build only to check that a planted stale async chunk disappears, and
the self-contained widget HTML test duplicated assertions the host
artifacts suite already makes per artifact. host-artifacts now plants
the stale chunk before its existing production build and asserts its
removal, and its HTML artifact loop keeps the inline script/style
presence checks, so mcp-transports drops both duplicates.

…e unit pool

The three finalization-polling tests exercise readFinalizedEvalRun with
in-memory fake clients and never touch a browser or server, yet they
lived in evals-real.e2e.test.ts and paid the serialized integration
pool for it. They move verbatim to evals-finalized-run.test.ts, which
the unit config picks up by default.

…nstead of the default integration pool

Every npm pack + clean-install suite (dev-workbench-packaging,
packed-consumer, packed-native-smoke, release-audit,
rsc-runtime-optional-packaging) leaves the serialized integration pool
and moves to a dedicated packedTestFiles list that the unit pool also
excludes. The three pack+install cases in public-api.test.ts split into
public-api-packed.test.ts so the cheap export and built-entrypoint
checks stay in the default loop. test:packed now lists the split file
plus rsc-runtime-optional-packaging and packed-native-smoke, keeping
per-PR CI coverage through the release-gates job's check:release run.

…service tests

Removes tests that only re-assert their own fixtures, string-match
documentation, or duplicate a sibling suite: the public-api config/type
re-export tests (the built-entrypoint test still pins subpath imports
and defineConfig identity), the packed-consumer regex self-test, the
manifest re-export identity test, the examples README string test, the
topology capture-command README pin, the audiobook tool-catalog and
CLI-receipt duplicates, the eval-cli duplicate of the eval-service run
path, the release-audit pack dry run (check:release runs pack:dry-run
directly), and one redundant JSON.stringify assertion in the native
smoke. Type-only contracts previously wrapped in vacuous runtime tests
(modern MCP transports, runtime provider binding/surface shapes) stay
as module-level @ts-expect-error checks. Deletes the orphaned
epoch-atomicity spike (production coverage lives in epoch-store and
dev-lock tests) and the micro-eval spot-check suite together with its
test:spot-check script and CI job; the examples:check step from that
job survives as its own examples-check job.

…-metadata rehash test

adapter-metadata.test.ts already rehashes every capability and schema
snapshot against its pinned provenance for all built-in targets, so the
host-adapters CLI-version test keeps only its unique assertions: the
observed CLI version pins, the redacted help text, and the codex
marketplace validator fixture.

…ifact writers into a serial pool

The parallel pool runs mkdtemp-fixture, ephemeral-port files on
min(4, cores/2) workers (still 1 on two-core CI, overridable via
AGENT_BUNDLE_INTEGRATION_MAX_WORKERS); the six files that rewrite shared
package dists stay on one worker in a chained serial config.

overview.e2e, mcp-app-real.e2e, and playground-real.e2e now use the shared
memoized buildWorkbench (which honors AGENT_BUNDLE_WORKBENCH_PREBUILT)
instead of local copies that rebuilt packages/workbench/dist up to ten
times per run; cli.test and the packed-release harness honor a new
AGENT_BUNDLE_PACKAGE_PREBUILT flag, with `test:integration` building the
whole workspace once up front. With those writes gone, the four files
move from the serial pool to the parallel one.

…o the eviction suite

The session accepts a test-only maximumRunHistory through its start
testing seam; the consolidated eviction test drives a five-run window
instead of fifty real invocations and asserts the production default is
still fifty.

…rallel workers

Concurrent Chrome + dev-server + rsbuild pairs contend for cores, so the
parallel config doubles timeScale for its workers and the e2e suites with
previously unscaled budgets now honor it (same mechanism CI already uses
on two-core runners).

Real in-process example builds can exceed the 5s rstest default when four
workers share the machine; explicit per-test timeouts are unaffected.

…gration-suite

# Conflicts:
#	packages/workbench/tests/overview.e2e.test.ts
#	packages/workbench/tests/runtime-playground.e2e.test.ts

…ration pool