Every engine exit is reported as "probably out of RAM", even when the engine's log names the cause #215

When the engine process ends during a request, the server prints death_note() (serve/server.py#L205-L221, used at #L856-L862). Unless the end of the log has the watchdog's issue #29 line, the note is always "The usual cause is running out of RAM ... Close other programs or use a smaller model". It does not look at the exit code or at the engine's own last log line.

We saw this on Windows 11 after the crash from #183 (exit code 1, the log ends with strata serve: cancelled) and after deliberately killing the engine mid-generation; neither had anything to do with memory. The note points users at RAM while the log already names the cause, and by its own text it cannot apply on Windows ("Windows slows down instead").

Reproduce (no GPU)

A stand-in engine that fails its first request the way #183's engine does, saved as fake_engine.py (executable) in the checkout root:

#!/usr/bin/env python3
import sys
print("READY 32768 stop", flush=True)
for line in sys.stdin:
    if line.startswith("GEN"):
        sys.stderr.write("strata serve: cancelled\n"); sys.stderr.flush()
        print("ERR cancelled", flush=True)
        sys.exit(1)

import sys, threading
sys.path.insert(0, "serve")
from server import StrataEngine

open("fake_engine.log", "w").close()
eng = StrataEngine("./fake_engine.py", [], log="fake_engine.log")
for attempt in (1, 2):
    try:
        for _ in eng.generate([1, 2, 3], 8, {}, threading.Event()):
            pass
    except Exception as e:
        print(f"request {attempt}: {type(e).__name__}: {e}")
print("log tail:", open("fake_engine.log").read().strip().splitlines()[-1])
print("death_note:", eng.death_note())

Output:

[strata] starting the engine: reading the model's weights ...
request 1: ValueError: cancelled
request 2: EngineDied: the engine stopped unexpectedly (exit code 1)
log tail: strata serve: cancelled
death_note: The usual cause is running out of RAM: Linux then ends the biggest program (check: sudo dmesg | grep -i -E 'killed process|out of memory'); Windows slows down instead. Close other programs or use a smaller model (Q2_0 / IQ2_XS).

Suggested

Give the RAM explanation only when the exit looks like the OS killed the engine (on Linux a negative return code, -9 for the OOM killer's SIGKILL), and otherwise print the exit code and the last strata ... line of the log, which death_note() already reads.