node-api: widen the measured slice, harden the harness, record the runs

Brings the node-api suite to an honest measurement and records the progression
35.8% → 43.2% → 50.4% → 60.1% → 65.0% → 72.3%. The latest run is 1,542 passes
out of 2,133 counting the denominator without skips.

That run is RED: 18 tests the ratchet expects to pass do not — 7 timeouts,
clustered in http and net, and 11 functional failures. The ratchet is
deliberately left as it was, so those 18 stay visible instead of being absorbed
into the baseline.

Harness

• Reap the whole process group on timeout. Node's core suite spawns children
  that outlive the test process and hold its stdout open, so killing the entry
  process alone left the output read hanging and the run stalled. The process now
  gets its own group and a timeout reaps the tree. A descendant that opened its
  own session still escapes the group kill, so the pipes are dropped after a
  two-second grace.
• Dereference the local install when staging the image. --elide usually
  points at the dist/current symlink, and the image build context must carry
  the real files; the staged copy now dereferences instead of building around a
  broken link.
• The cgroup cage is optional. Probing systemd-run no longer fails the run:
  Bun.spawnSync throws for a missing executable rather than reporting an exit
  code, so inside the harness container the whole suite died before its first
  test. An absent systemd-run now means "run uncaged" and the container's own
  limits apply. A test covers that branch: it strips the PATH of the child it
  spawns rather than mutating process.env, because the search path a process
  uses is the one it started with.
• Registry: the test expects the configured jtreg case timeout rather than
  the previous default.

Test selection

• Coverage widened with the missing subsystem groups: child_process, cluster,
  dgram, domain, http, https, net, os, perf_hooks, querystring,
  readline, tty, v8, zlib and more, so the suite measures the public Node
  surface rather than a sparse slice.
• Memory bound per test: -XX:MaxHeapSize=4g. The suite contains cases whose
  only brake is stream backpressure (test-zlib-brotli-16GB.js decompresses
  16 GB); unbounded, such a case exhausts the host instead of failing.
• New static expectations file expectations/node-api.toml carries the skips with
  a reason each: the cluster group is out of project scope for now (one glob, so
  the decision stays visible and reversible), and the node:vfs tests cover an
  experimental Node subsystem Elide does not implement.

Ratchet and reports

• The ratchet holds 857 known failures. It is deliberately not retaken against
  the newest run, so the 18 regressions above stay reported rather than absorbed.
• Six runs are recorded. The latest carries its full report body — the impact
  map is the working document for the next batch of fixes. The earlier ones
  keep only the raw results, counts and change summary: that is what the trend,
  the run database (ingestAll) and the next run's diff actually read, while the
  rendered bodies are superseded as soon as a newer run lands.
• BRIEFING-node-api.md is rewritten around the current measurement: the failure
  map by subsystem with the hang subset broken out, what the static expectations
  skip and why, the ranked work items that follow, and run instructions matching
  the current selection.

Checks

bun --cwd harness test — 172/172 pass. bun run typecheck — clean.