Harden flaky native process-lifecycle tests for slow CI

[dakra]

Three native tests flake on the heavily-loaded GitHub runner, which runs make -j$(nproc) test-native — spawning a real PTY child per test file in parallel:

• ghostel-test-exec-delete-process-kills-sighup-ignoring-child
• ghostel-test-exec-kill-buffer-leaves-sighup-ignoring-child-live
• ghostel-test-compile-kill-compilation-finds-live-buffer

Two distinct causes:

1. Regression in kill-compilation-finds-live-buffer

The readiness handshake added in 36d9510 ("Harden flaky tests") was dropped in 9e8460a ("Make ghostel buffers read-only"). Without it, kill-compilation (SIGINT) races process startup: the signal can land before /bin/sh -c "exec /bin/sleep 30" has exec'd sleep / the PTY foreground process group has settled, so finalize may not happen within the wait. Restore the GHOSTEL_KILL_READY marker and wait for it before signalling.

2. Tight 5s startup waits under parallel CI load

The two exec SIGHUP tests are logically sound (verified both PTY backends: kill-buffer → SIGHUP which the child traps, delete-process → SIGKILL), but rely on a hard-coded 5s startup wait. That's too tight under -j$(nproc) load, made worse now that the native PTY backend runs them twice via with-pty-matrix. The same 5s recurs across osc/exec/spawn/shell/keys native tests.

Fix centrally: add ghostel-test--timeout-scale (1 locally, 4 when CI is set, overridable via GHOSTEL_TEST_TIMEOUT_SCALE) and apply it in ghostel-test--wait-for / ghostel-test--wait-until. Every polling wait routes through these two, so the whole suite gets headroom in one place. The helpers early-exit on their predicate, so the larger ceiling costs nothing on passing runs. The scale gates on a positive numeric parse, so an empty or malformed override falls through to the CI/default logic instead of zeroing every timeout.

The unscaled fixed-window/negative assertions (accept-process-output proc 0.2, etc.) are deliberately left alone — scaling a "nothing should happen within this window" guard would change its meaning.

Verification

• Three named tests pass locally (scale=1) and under simulated CI (CI=true, scale=4), both fast.
• Scale resolution checked across unset / CI / empty / garbage / zero / valid / fractional overrides.
• make -j8 all → exit 0, all suites green. The changed helper is a prerequisite of every test stamp, so the full elisp + native suite re-ran under the new code.