fix: stop phantom CoS agents from blocking the update page forever

[atomantic]

Summary

The Apps → Update page sat on "Update paused — 4 CoS agents running" with an
empty agent list underneath, and no way out: updating restarts PortOS, and that
restart was the only thing that would have cleared the count.

The four agents were real processes — codex TUIs whose PRs had merged hours
earlier — that nobody could kill.

Root cause, in four layers

1. node-pty's Windows backend throws Signals not supported on windows. for
   any signal argument. Every TUI kill that passed one was a logged no-op:
   ❌ TUI termination relay failed: Signals not supported on windows. The PTY
   survived, held its worktree locked (EBUSY on cleanup), and stayed in the
   CoS Runner's active set. This hit every TUI termination path — the runner's
   /terminate, /kill, /pause, /terminate-all, the server's
   post-finalize tui:kill relay — plus stopRun on a TUI prompt run through
   the vendored aiToolkit runner, and terminate/kill of a direct-mode TUI agent.
2. The runner kept advertising the survivors in GET /agents.
3. syncRunnerAgents adopted whatever the runner advertised, minting phantom
   runnerAgents entries that nothing ever pruned.
4. The update gate counted map entries, so the phantoms blocked updates forever.

Changes

Each layer is fixed where it belongs.

killProcessTree (server/lib/bufferedSpawn.js, mirrored in the
self-contained aiToolkit/runner.js) already knew a node-pty handle from a
spawned child — it just forgot the second half of the rule. A non-ChildProcess
killable is now offered the caller's signal first and falls back to a bare
kill() only when the handle rejects it, so node-pty kills land while the CoS
runner's signal-forwarding socket proxy keeps its force-kill. It takes an
injected platform flag so both Windows branches are covered on a Linux runner.

The CoS runner routes every kill through that helper and shares one
force-kill escalation (forceKill.js) instead of four hand-rolled copies. The
escalation drops the agent from the active set, so a process that outlives its
own SIGKILL can no longer be advertised as a live run.

syncRunnerAgents adopts only agents PortOS still holds a live durable
record for, and the orphan sweep gained the reverse pass that retires
runnerAgents entries with nothing behind them. A failed record read stays
distinct from an absent record — one leaves live state alone, the other
retires it.

The update gate counts only tracked ids whose durable record is still live,
so a run PortOS already finalized cannot hold a restart back. It fails
closed: the records are read through a non-defaulting
readAgentsStateForSafetyCheck (sibling of the Persistent Mind gate's own
trusted read), because loadState substitutes an empty state for a corrupt
file — which would have handed the gate a confident zero and restarted PortOS
out from under live agents.

Defects the fix exposed, closed in the second commit

• armForceKill refused to arm over a timer /pause had left behind, so
  terminating a paused agent skipped its escalation entirely.
• The runner's TUI exit handler had no paused guard (its CLI sibling does), so
  a paused TUI was finalized as failed and its task retired — unreachable
  while the kill was throwing, live the moment it landed.
• A kill relayed for a paused agent reaped the durable record a later resume
  reads.

Test plan

• server/cos-runner/forceKill.test.js — new; exercises the escalation for real
  (drop-on-grace, re-arm over a stale pause timer, handle clearing, dropState
  scoping, throw containment). The stale-timer case fails against the previous
  implementation, which is how that regression was caught.
• server/lib/bufferedSpawn.test.js — the pty cases now drive an injected
  platform instead of if (!IS_WIN32) return, so they assert on every runner.
• filterLiveAgentIds / readAgentRecordOrUnreadable unit tests, the
  update-route gate tests, the runner-sync adoption guard, and the orphan
  sweep's reverse prune all pin their contracts and fail if reverted.
• Test mocks now import the real isLiveAgentRecord rather than
  re-implementing it; fixtures standing in for a spawned child carry the
  ChildProcess prototype so they exercise the spawned-child branch.
• Full server suite: 493 affected suites green (11 912 tests). The 15 remaining
  failures on this machine are pre-existing and environmental (missing
  playwright-core, no Python).

Follow-up filed as #5622: the zombie and orphan sweeps still treat the runner's
agent list as proof of liveness.