fix(cli): run --force creates a nested session inside a pty (parity R1 #4)

[myobie]

Parity drive R1 — node reference spec + run --force fix (#4)

Locks node pty's observed behavior so the Rust port (pty-rust) can match it, and
fixes the one R1 divergence CoS ruled on.

The #4 fix (behavior change)

pty run --force from inside a pty session used to treat --force as a
no-op: it still ran the command directly in-place and created no session.
That contradicted:

• pty's own run --help — "Create even from inside another pty session"
• attach/restart --force symmetry (those bypass the nesting guard)
• the pty-rust port

CoS ruling (decision a, 2026-07-21): fix the node code so run --force
bypasses run's nesting guard and creates a real (nested) session. --force
now behaves like -d w.r.t. the guard — both fall through to session creation
(-d → detached/background, --force → interactive/attached).

Changes

• src/cli.ts — add && !force to the run nesting guard so --force
  falls through to cmdRun (create + attach a nested session). The inner -a
  "refuse when target running" sub-check is now reached only on the no-force
  path, so its redundant !force is dropped. Comments updated.
• tests/parity-node-reference.test.ts (new) — the R1 reference spec: 6
  tests over the 4 divergences (post-exit peek preserves final screen, plain
  peek keeps the trailing cursor cell, --seq pacing incl. ms rounding, and
  posix_spawnp fails on Node.js v25 #4). posix_spawnp fails on Node.js v25 #4 now asserts run --force CREATES a session via async-spawn +
  registry poll (deterministic; no foreground-attach hang).
• tests/nesting-prevention.test.ts — the old "--force overrides when
  nested + target running" test encoded the removed no-op behavior (and hung
  ~10s on a blocking in-place cat). Rewritten to deterministically assert the
  new bypass-and-create behavior. ⚠️ reviewer note: this is a necessitated
  test change, slightly beyond "the 6 tests + posix_spawnp fails on Node.js v25 #4 fix" — the code fix inverts
  that test's premise, so leaving it would ship a test with a now-false comment.

Verification

• Full suite: 1383 passed, 21 skipped (pre-existing env-gated skipIf), 0 failed.
• New parity posix_spawnp fails on Node.js v25 #4 test runs in ~317ms (no hang).
• dist/ is gitignored; rebuilt locally before running (tests exercise dist/cli.js).

Coordination

pty-rust is already in lockstep at CREATES (flipped + pushed, 167 green).
Node was the outlier; this brings code == docs == rust. R2 shared fixtures next.

/cc CoS for review + merge (author does not self-merge).