Sylpheed/docs/re/upstream-baseline.md

A stock-upstream baseline runs Stage 02 crash-free

Status: ✅ CONFIRMED — upstream canary_experimental + only the pad driver and the threading fix reaches Stage 02 flight with zero crash dumps, where the instrumented branch produces 384 at the briefing and never reaches flight. 🔴 pure upstream on its own is less stable, not more — it still has the lost-resume race. 🔴 the REMAINING OB address is not found: the first candidate was refuted.

Why this was tried

The instrumented branch stops at the Stage 02 briefing under a storm of 0x82307128 (title-crash-stl-tree.md). The question "is that crash ours or upstream's?" is answerable by building upstream clean.

Building it, and two tooling gaps found on the way

git fetch upstream works from this container. The merge base is 82 commits behind upstream/canary_experimental (a5a18f5c7); our branch carries 50 of its own.

• version.h is never generated. The build fails on trace_writer.cc:17: fatal error: 'version.h' file not found. Upstream's xenia-build.py writes it from git HEAD; the container's build-canary wrapper does not invoke it, and our tree only builds because a stale copy from an old sylpheed-re build sits in the build directory. Regenerated by hand in exactly the format that script emits.
• build-canary reports success on a failed build. The harness recorded "completed (exit code 0)" while ninja had stopped with 1 error generated. Only the missing binary gave it away.

Pure upstream cannot be driven — and is not more stable

--hid=file is ours (d15c8cfab and two follow-ups), and there is no uinput in this container, so a genuinely pure binary boots but cannot be scripted past the title. The honest baseline is therefore upstream + those three commits.

That baseline reproduced the lost resume on its second boot: black screen, guest alive at 522 % CPU, the newest host thread at 00:00:00 CPU time, its guest handle making zero kernel calls — the exact signature from canary-scripted-input-traps.md. Reading upstream/canary_experimental:src/xenia/base/threading_posix.cc confirms the two-lock-scope publish is still there. So "pure upstream" is not a stable state: it contains a real bug this project had already fixed.

✅ Upstream + the threading fix: Stage 02, in flight, zero crashes

Adding only a60fe7d11 on top — still none of the RE instrumentation — launch_mission.sh drove boot → title → LOAD GAME → slot 01 → READY ROOM → TAKE OFF → flight:

TIME 05:03.75, wingman Rhino 103 / Ellen Bernstein, radio dialogue, asteroids, tracer fire — and REMAINING OB 020 with OB badges on the objective targets. The run ended in a legitimate GAME OVER (the unattended craft was shot down, exactly as autopilot-memory-driven.md warns) — not a crash.

Crash dumps for the entire run: 0. Against 384 on the instrumented branch at the briefing alone.

So the mission crash is not upstream's. It is either in our 50 commits or fixed among upstream's 82 — this run does not separate those two, and saying which would need a bisect.

🔴 REMAINING OB in RAM: first candidate refuted

The counter is visible and moves (020 → 019 → 018 → 017), so it is scannable.

• A three-snapshot big-endian u32 filter (19 → 19 → 18) left 0 survivors.
• A direct differential across widths over the 19→18 transition gave exactly one aligned BE u32: va 0xbc22e83c (the u16 hit at …83e is its low half).
• Refuted. Read live moments later it held 26 while the HUD showed 017. It is an unrelated counter that happened to step 19→18 in the same window.

That the u32 filter came back empty is itself a result: the counter is not a plain persistent big-endian word. Candidates worth trying next are a per-digit representation, a value recomputed each frame from an entity list (in which case the list is the real target), or a field inside a structure that moves.

Not settled

• 🔴 Which side owns the mission crash — our 50 commits or upstream's 82. A bisect would say; nothing cheaper will.
• 🔴 The REMAINING OB address.
• ⚠️ wait_flight.sh reported NEVER REACHED FLIGHT while the game was plainly in flight — its detector is wrong, and screen_id.py got it right.
• ⚠️ Any future in-mission scan needs the pilot flying or a safe holding pattern: an idle craft is dead in about a minute.