vwemirror/BT411
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
1b6d247e51e775d43d2bb823451da9d2c542e17e
BT411/docs/MISSION_END_AUDIT_20260811.md
T

59 KiB
Raw Blame History

The mission-end / exit-hang audit -- full findings (2026-08-11, ticket #163)

Four-agent audit of the round-end freeze + hang-on-exit family. VERDICT: the mission-end ladder is CLEAN on every machine (StopMission never even fires in steam fields -- each pod self-ends on its local clock); the bug lives in the unlogged exit window: bare ExitProcess with no SteamAPI_Shutdown anywhere in the codebase -> intermittent DLL-detach deadlock -> the dead process's fullscreen frame squats on screen while its menu child hangs in SteamAPI_Init against the wedged sibling (field-captured: ZEUS 22:12:38). Fix shipped same day: BTSteamNet_ShutdownAll + ordered exit (flush, steam down BEFORE child spawn, hide windows, TerminateProcess) at every exit path. Side-yield: #156's fade root cause (POVStartEndRenderable ctor is an empty stub, btstubs.cpp:339) + the complete binary end-of-mission ladder.

#163 — Task 2: The end-of-mission ladder as built, every stall-capable site flagged

Stall-capable sites are tagged [S#]. Evidence tiers: [T0]=our engine source read directly; claims about runtime behavior of third-party code marked [T4].

Stage A — the HOST marshal (game/glass/btl4console.cpp)

Arming. btl4main.cpp:1322-1330 — if BT_FE_EGG is set, BTLocalConsole_Start(fe_egg, BT_FE_PODS, atoi(BT_FE_SECS)) runs. For Steam hosting the menu sets BT_STEAM_NET=1, BT_FE_MYFAKE, BT_FE_STEAMMAP and falls into the marshal-armed default (btl4main.cpp:1284-1296). BTLocalConsole_Start (btl4console.cpp:588-606) copies the args into marshalState and CreateThread(MarshalThread); the handle is closed immediately (:604) — the marshal thread is never joined by anything.

MarshalThread (btl4console.cpp:434-586), step by step:

  1. Read the egg (:440-446). Failure → MarshalLog("cannot read egg...") and the thread just returns — the pods sit at WaitingForEgg forever [S0] (not tonight's shape, but a stall).
  2. Connect every pod in BT_FE_PODS order, self first, sequentially (:452-482): ConnectWithRetry(host, port, 60) (:388-432). The Steam branch calls BTSteamNet_Connect which blocks up to 30 s per attempt (L4STEAMNET.cpp:595-609) — worst case 60×~31 s ≈ 31 min per pod [S1]. Any pod that never answers → BTFE_RelaunchSelfAndExit("") — the whole host process aborts the start (:473-476).
  3. Egg → every pod (:492-497) via SendEggChunks (:334-363). The Winsock branch uses blocking send() — the marshal's raw sockets are never set FIONBIO (only TCP_NODELAY, :423-425) [S2]. Errors surface only as a -1 chunk count in the log; no retry.
  4. Sleep(20s); RunMission #1 to all; Sleep(4s); RunMission #2 to all (:500-512). Send results ignored.
  5. The mission clock (:519-570): mission_end = GetTickCount() + missionSeconds*1000; loop drains and discards pod→console traffic (select() for real sockets, a nonblocking BTSteamNet_Recv pass for Steam pseudo-sockets :559-569). Note: a Steam connection that dies mid-mission is invisible here — closedByPeer makes Recv return 0/err and the loop just moves on; nothing logs or repairs it.
  6. StopMission at expiry (:572-576): 
    MarshalLog("mission clock expired -- StopMission to all pods");
for (int p = 0; p < pod_count; ++p)
    SendApplicationMessage(pods[p], StopMissionMessageID);
    SendApplicationMessage (:369-382) writes one 28-byte packet (header clientID=4/ApplicationClientID, msg id 6, ReliableFlag). One shot per pod. No retry. No acknowledge. The return value is ignored. There is no per-pod success/failure log — only the single "clock expired" line [S3 — the silent-drop site]. Over the Steam seam, BTSteamNet_Send (L4STEAMNET.cpp:424-442) returns SOCKET_ERROR if the connection is closedByPeer (WSAECONNRESET) or if SendMessageToConnection fails (WSAEWOULDBLOCK) — all swallowed. Over raw TCP the blocking send could also wedge the loop between pods so later pods in podList never get their stop [S4] (needs a full send buffer; unlikely for 28 bytes).
  7. Sleep(MissionEndGraceSeconds=8s) (:577), MarshalClose every pod (:578-581), BTFE_RelaunchSelfAndExit("") (:583-584) → Stage D. During the 8 s grace the marshal no longer drains pod traffic (pods push an immediate console score update in EndingMission — btplayer.cpp:1361-1366).
  8. Host-side race [S5]: the host's own pod (self, first in podList) receives StopMission over its loop link, finishes its fade in ~3-4 s, and the MAIN thread reaches the BT_FE_LOOP relaunch (btl4main.cpp:1669-1674) → ExitProcess — while the marshal thread is still inside its 8 s grace sleep. Whichever thread reaches ExitProcess(0) first kills the other mid-flight; if the main thread is slow (long Shutdown) both can pass CreateProcessW → two menu children. Nothing serializes the two relaunch paths.

Stage B — the receiving pod: StopMission → RunMissions returns

Delivery. The 28-byte packet is routed by networkManager->RoutePacket(), pumped at least once per frame from Application::ExecuteBackgroundTask (APP.cpp:771-781; in non-Running states it routes greedily). Steam-carried console bytes are drained lazily by the seam's recv (L4STEAMNET pump is called from the seam entries).

Handler chain: BTL4Application::StopMissionMessageHandler (game/reconstructed/btl4app.cpp:765-781, @004d3a94 — plasma display off) → L4Application:: (L4APP.cpp:693-732 — PilotIllumination(True) + a LightsOut self-post at +30 s) → Application:: (APP.cpp:1706-1778):

  • gBTMissionStoppedByConsole = 1 (:1723-1726); if state==RunningMission: gBTRoundCompleted=1, launch pends voided (:1733-1739).
  • switch (:1746-1777): StoppingMission → ignore; EndingMission/AbortingMission → Stop(); default → SetState(EndingMission), networkManager->Mode(ReliableMode) (:1762), dispatch Player::MissionEndingMessage to GetMissionPlayer() (:1759-1769); if the mission player is NULL → Stop() immediately (:1771-1774).

The fade. Player::MissionEndingMessageHandler (engine/MUNGA/PLAYER.cpp:106-113): fadeTimeRemaining = 3.0f; ForceUpdate(); SetSimulationState(MissionEndingState). BTPlayer has NO override — MESSAGE_ENTRY(BTPlayer, MissionEnding) (btplayer.cpp:355) resolves to the inherited Player handler; the "fade ForceUpdate site" named in the task brief is PLAYER.cpp:110.

EVERY condition required to progress from EndingMission to Stop() — the countdown lives ONLY in Player::ManageApplicationStatus (PLAYER.cpp:408-476; the MissionEndingState arm :444-472 counts fadeTimeRemaining -= time_slice and, at ≤0 with app state EndingMission, dispatches StopMissionMessage(NullExitCodeID) back at the application :451-456):

  1. A mission player must exist (else Stop() already ran — faster, not a stall).
  2. ManageApplicationStatus is reached ONLY from the player's active Performance: Player::PlayerSimulation (PLAYER.cpp:505, called from BTPlayer::PlayerSimulation btplayer.cpp:1329) or CameraShipSimulation (PLAYER.cpp:486). If the mission player's Performance is HuntForDropZone (PLAYER.cpp:768 — no MAS call) or DoNothingOnce (set at PLAYER.cpp:380 once a dropzone answered the initial hunt; DoNothingOnce = NeverExecute(), SIMULATE.cpp:484-488 — the entity STOPS EXECUTING ENTIRELY), the fade NEVER counts down and the app sits in EndingMission forever [S6 — hard in-engine stall]. This window exists from mission start until DropZoneReply creates the vehicle and sets PlayerSimulation + AlwaysExecute() (btplayer.cpp:1771/1780/1784). After that, the reset-based respawn keeps the same Performance through death, so a dead-at-mission-end pilot still fades normally.
  3. Fade 3 s of simulated time must elapse; then the re-dispatched StopMission lands in the EndingMission case → Stop() (APP.cpp:1751-1754) → executeFrames = False + SetState(StoppingMission) (APP.cpp:818-820).
  • There are no queue-drain gates, no all-entities gates, no peer gates, and no other timers in this leg.

Backstop worth knowing: the solo-clock shim (APP.cpp:674-695) is ACTIVE in Steam matches — its only gate is getenv("BT_RELAY") == 0, and glass/steam mode does not set BT_RELAY; the glass egg writes length= (btl4fe.cpp:277) → Mission::gameLength > 0 (MISSION.cpp:336-337). So a pod that never receives the marshal's StopMission still self-posts one when its own clock (started at RunningMission entry, APP.cpp:1559) expires. A lost StopMission alone therefore should not strand a pod in RunningMission [T0 code-read; runtime unverified].

RunMissions returning (engine/MUNGA/APPMGR.cpp:39-273): next frame ExecuteForeground returns False (APP.cpp:516-519) → APPMGR calls application->Shutdown() (APPMGR.cpp:77-91). Application::Shutdown (APP.cpp:828-927): gauge/video/audio Shutdown(), viewpoint delete, interestManager->Shutdown(), hostManager->Shutdown(), networkManager->Shutdown() — L4NetworkManager::Shutdown (L4NET.CPP:1296-1411) closes every non-console host via CloseConnection = shutdown(SD_BOTH) + BTNetClose (L4NET.CPP:4658-4659), all nonblocking (engine sockets are FIONBIO throughout, e.g. L4NET.CPP:4554-4558) — deletes the mission, and returns False (APP.cpp:925). App is moved to the ended list; the next loop pass finds no apps → Terminate() (APP.cpp:934+ — deletes renderers/managers; D3D/DirectSound teardown lives here [S7, T4]) → delete application → RunMissions returns (APPMGR.cpp:128-138). Alternate exit: WM_QUIT (window destroyed) → Terminate + return (APPMGR.cpp:102-115).

Stage C — after RunMissions returns (game/btl4main.cpp:1618-1684)

  • :1619 "[boot] RunMissions returned (mission loop exited)." — the key forensic marker separating a Stage-B stall from a Stage-C/D hang.
  • :1623-1626 BTProjectilesClearAll() — static pool scrub, memory-only.
  • :1632-1635 BTRelayUploadMatchLog() — relay-only: gated on s_relayGameAddrCached (L4NET.CPP:707-708), a no-op on Steam nights. When it does run: bounded 3 s nonblocking connect (L4NET.CPP:738-753), but then an unbounded blocking SendAll (L4NET.CPP:770) [S8, relay only].
  • :1645-1660 BT_RELAY rejoin — not taken under Steam.
  • :1669-1674 BT_FE_LOOP set (it always is on menu-launched missions, btl4main.cpp:1248) → BTFE_RelaunchSelfAndExit("").

Stage D — the exit/relaunch choke point (btl4console.cpp:156-276)

  1. gBTUserRequestedExit (WM_CLOSE stamp, btl4main.cpp:164-172) → ExitProcess(0) directly (:171-181).
  2. Storm damper: generation <15 s → Sleep(5000) (:193-202).
  3. Menu relaunch clears BT_FE_EGG/PODS/SECS/LOOP from the env (:211-217).
  4. CreateProcessW of the menu child (:259-274). A CreateProcessW failure is completely silent — no log, no retry — and control still falls to ExitProcess: window closes, no menu appears [S9].
  5. ExitProcess(0) (:275) — the exit-hang site [S10]. ExitProcess terminates all other threads (the marshal, or the main thread if the marshal got here first) then runs DLL_PROCESS_DETACH serially under the loader lock. SteamAPI_Shutdown() is never called anywhere in the tree (grep hits only the SDK header, extern/steamworks_sdk_164/public/steam/steam_api.h:107) — so steam_api64/steamclient are detached with live SDR connections and their service threads hard-terminated; a detach-time deadlock there is the standing hypothesis for "BTL4 hang on exit, hosting" [T4 — inference, matches the host being the machine that hosts the most Steam connections]. D3D/driver DLL detach is the sibling candidate [T4].
  6. Because every glass-path exit is ExitProcess, CRT atexit/static dtors are skipped — the RIO dtor's bounded 5 s flush (L4RIO.cpp:1160-1184) and PCSerialPacket::ShutdownRxThread's unbounded WaitForSingleObject(hRxThread, INFINITE) (L4PCSPAK.cpp:204) [S11] only run on the plain return Exit_Code path (btl4main.cpp:1683, taken only when BT_FE_LOOP is unset) — and only in serial-RIO builds; glass uses PadRIO which owns no thread (no CreateThread in L4PADRIO.cpp).

Stage E — is there any all-pods handshake at mission end? NO.

  • Start side has three gates: the egg acknowledge (AcknowledgeEggFileMessage, L4NET.H:276/424-433), the READY notify (APP.cpp:1466-1470), and the connection gate (L4NET.CPP:1957 comment). The end has none: no end-acknowledge message exists (grep across engine/), the marshal reads nothing back after StopMission, and pod teardown never waits on a peer.
  • HostDisconnected during EndingMission: GameMachineHostType arm just decrements the count and closes (L4NET.CPP:1672-1677) — no state change, no wait. Console loss mid-mission in glass/steam (non-relay, scene presented): the ConsoleHostType arm recreates the console listener and continues the mission (L4NET.CPP:1686-1753) — the relay-mode 15 s graceful self-stop (RelayGameDown, L4NET.CPP:2458-2472) does NOT apply outside BT_RELAY.

Cross-cutting observations relevant to the incidents

  • Thread-safety hole [S12]: BTSteamNet_Pump (SteamAPI_RunCallbacks + unlocked connections[].ring writes, L4STEAMNET.cpp:244-306) is invoked from BOTH the marshal thread (via BTSteamNet_Send/Recv/Connect — btl4console.cpp:119-124, 564) and the game thread (seam entries; L4STEAMNET.h:34 documents "game thread"). No lock anywhere. Concurrent RunCallbacks + ring index races are possible on the HOST only (the only process with a marshal) [T0 for the code paths; consequences T4].
  • Frozen-view interpretation [inference]: a stalled EndingMission does NOT freeze the picture — ExecuteForeground has no EndingMission early-out, the update manager keeps simulating and the renderers keep presenting (APP.cpp:560-565 only overlays pre-run states). A literally static final frame with no menu means the frame loop stopped while the window lived: i.e., RunMissions returned (or Terminate/Shutdown hung) and the process then wedged in Stage C/D — most plausibly at ExitProcess [S10] or silent CreateProcessW failure [S9]. Log triage for the frozen machines: "[boot] RunMissions returned" present → Stage C/D hang; absent but StopMission/plasma-off lines present → Stage B stall ([S6]/[S7]); no stop lines at all → the marshal's send was silently dropped ([S3]) and the solo-clock backstop should then have fired ~launch-skew seconds later — if it didn't, check length= in the session egg.

Stall-site index

# Site File:line Bound
S0 egg unreadable → marshal thread exits, pods stranded btl4console.cpp:440-446 forever
S1 ConnectWithRetry, Steam branch btl4console.cpp:388-408 + L4STEAMNET.cpp:595-609 ~31 min/pod
S2 blocking egg/TCP sends (marshal sockets never nonblocking) btl4console.cpp:334-363, 409-431 OS send timeout
S3 StopMission one-shot, errors swallowed, no per-pod log btl4console.cpp:572-576, 369-382; L4STEAMNET.cpp:424-442 n/a (silent drop)
S4 blocking StopMission send can starve later pods btl4console.cpp:573-576 TCP buffer dependent
S5 main-thread relaunch races marshal's 8 s grace; dual relaunch btl4main.cpp:1669-1674 vs btl4console.cpp:577-584 race
S6 fade counts only in PlayerSimulation; DoNothingOnce=NeverExecute strands EndingMission PLAYER.cpp:380, 444-472; SIMULATE.cpp:484-488 forever
S7 Terminate/Shutdown renderer teardown (D3D/DSound) APP.cpp:838-865, 944-973 unbounded [T4]
S8 matchlog upload blocking SendAll (relay only) L4NET.CPP:770 TCP dependent
S9 CreateProcessW failure silent, no menu btl4console.cpp:259-274 n/a
S10 ExitProcess w/ live SteamAPI (no SteamAPI_Shutdown anywhere), DLL detach btl4console.cpp:275; steam_api.h:107 unbounded [T4]
S11 PCSerialPacket join INFINITE (static-dtor path only, non-glass) L4PCSPAK.cpp:204 unbounded
S12 unsynchronized two-thread Steam pump on the host L4STEAMNET.cpp:244-306; btl4console.cpp:119-124 corruption risk

==============================================================================

Task 4 — Steam console-link delivery of StopMission (ticket #163)

1. The exact delivery chain for StopMission under Steam

Dial timing — at GO, once, never again. The host's mission process arms the marshal thread at boot (game/btl4main.cpp:1322-1329 → BTLocalConsole_Start, game/glass/btl4console.cpp:588-606). MarshalThread dials EVERY podList entry up front, serially, before anything else: ConnectWithRetry(host, port, 60) at btl4console.cpp:470. For a fake-IP entry the Steam branch (btl4console.cpp:391-408) calls BTSteamNet_Connect up to 60 times, 1 s apart; each attempt itself blocks up to 30 s in the SDR handshake wait (engine/MUNGA_L4/L4STEAMNET.cpp:595-609) — worst case ~31 min of near-silence before the abort at btl4console.cpp:473-476 (which aborts the mission start for EVERYONE via relaunch). [T1 — read from our own source]

Token resolution. BTSteamNet_Connect maps the token IP to a SteamID64 via the tokens[] table (L4STEAMNET.cpp:556-568), loaded ONCE per process at BTSteamNet_Install from env BT_FE_STEAMMAP (L4STEAMNET.cpp:354-384); port 1501 selects P2P virtual channel 0 = console (L4STEAMNET.cpp:569, TokenConsolePort at :67). The connection is a ConnectP2P handle wrapped in a pseudo-SOCKET (0x5EA0xxxx, L4STEAMNET.cpp:93-97).

Send path — RunMission and StopMission ride the SAME connection. Egg chunks (SendEggChunks, btl4console.cpp:334-363), RunMission #1 (+20 s), RunMission #2 (+4 s), then the marshal holds the clock (:519-570) draining pod→console bytes, then StopMission at clock expiry (:572-576), 8 s grace, then MarshalClose all pods (:577-581) and relaunch (:583-584). There is no re-dial, no ack, no health check anywhere — StopMission depends on the connection dialed at GO still being alive ~10 min later. Wire message: 28-byte packet, clientID=4 (ApplicationClientID), messageID=6 (StopMissionMessageID, btl4console.cpp:75, per APP.h:383), sent via BTSteamNet_Send → SendMessageToConnection(..., k_nSteamNetworkingSend_ReliableNoNagle) (L4STEAMNET.cpp:433-435).

2. Every silent-failure point in that chain

  1. Send results are discarded. SendApplicationMessage returns -1 on failure but the return is ignored at all three call sites (btl4console.cpp:503, 509, 575). A failed StopMission produces NO log line; the marshal logs only the aggregate "StopMission to all pods" (:572) BEFORE sending. [T1]
  2. Send-on-dead-connection is an error only if the death was already noticed. BTSteamNet_Send returns SOCKET_ERROR/WSAECONNRESET only when closedByPeer was set by the status callback (L4STEAMNET.cpp:428-431), and SOCKET_ERROR/WSAEWOULDBLOCK when SendMessageToConnection fails (:436-440) — both swallowed by (1). Worse, a reliable send on a dying-but-not-yet-flagged connection returns k_EResultOK, buffers the bytes, and they are dropped when the connection is later closed — undetectable even if (1) were fixed. [T4 — SDK semantics, not exercised in a bench]
  3. The marshal never learns a console link died mid-match. The clock-hold drain loop treats Steam recv()==0 (dead: L4STEAMNET.cpp:459-462) identically to -1/no-data (while (BTSteamNet_Recv(...) > 0), btl4console.cpp:564); the real-socket path likewise ignores recv==0 (:550). No close callback reaches the marshal; the only trace is the engine-side [steamnet] connection N closed (…) line printed from the status callback during a pump (L4STEAMNET.cpp:219-231). No reconnect logic exists. [T1]
  4. Mesh/Steam pods have NO console-loss StopMission fallback. The 15 s self-stop on console loss is RELAY-ONLY (RelayGameDown, engine/MUNGA_L4/L4NET.CPP:2456-2472). In mesh/steam mode a ConsoleHostType disconnect just destroys and re-listens the console host (L4NET.CPP:1686-1752, CreateConsoleHost at :885-1023) — a pod that loses its console link mid-match plays a forever-mission: peers time out and freeze as stale replicants — literally "a frozen view of how the game ended" with no menu transition. [T1 code + T2 field, below]
  5. Pump concurrency hazard (host only). BTSteamNet_Pump mutates the global connections[]/rings and runs SteamAPI_RunCallbacks with zero locking (L4STEAMNET.cpp:244-306), and on the HOST it is called concurrently from the marshal thread (via Send/Recv/Connect, e.g. btl4console.cpp:564) and the main game thread (via the L4NET seam, L4NET.CPP:48-96). Only the host has this exposure. [T1 for the race's existence; T3 for consequences]
  6. Accept is channel-blind. BTSteamNet_Accept ignores which listener polled and dequeues either channel (L4STEAMNET.cpp:476-501); the console-host accept path trusts "whatever connects to a console host" (L4NET.CPP:3876-3894). Benign so far, but a game-channel dial can be adopted as the console. [T1, no field hits]
  7. Second-round state: the connection table is per-process (fresh each round — every round is a new process), and mission children get fresh BT_FE_MYFAKE/BT_FE_STEAMMAP at GO (btl4main.cpp:1279-1287). But a MENU relaunch clears only BT_FE_EGG/PODS/SECS/LOOP (btl4console.cpp:211-217) — menu children carry the PREVIOUS round's token map (visible in the field: relaunched menus log "N roster token(s) incl. self" at boot). The real cross-round asymmetry found is zombie processes: a round-N process wedged at exit stays Steam-online into round N+1 (see acaci's ghost lobby seat, below). [T1/T2]

Receiving side (pod). CreateConsoleHost listens on 1501 (L4NET.CPP:962-993) — under Steam the marshal's dial arrives via the P2P channel-0 listen socket, is accepted inside the status callback (L4STEAMNET.cpp:189-203), queued, and dequeued by the console host's BTNetAccept poll (L4NET.CPP:3814-3821); the console host goes OnLine (:1593-1595, logs "Connected to ConsoleHost at <token .1>:1501"). Marshal-side death is detected only by BTNetRecv returning 0 (or WSAECONNRESET→0) during CheckBuffers (L4NET.CPP:3989-4019). At mission end the pod does NOT close the console link — teardown explicitly keeps the console host (L4NET.CPP:1367-1370); the MARSHAL closes it 8 s after StopMission (btl4console.cpp:577-581), which the pod handles as a normal console disconnect + re-listen. So there is no close-while-the-other-side-needs-it on the pod side. [T1]

3. Field evidence (scratchpad/night15, six logs; frozen = santo/MS-FIREFLY, acaci/ZEUS; Lynx has NO log)

StopMission WAS delivered to every logged pod in both incident rounds. Every mission session on santo ends with the full clean tail [boot] RunMissions returned + [fe] mission over -- relaunching the menu (14/14, e.g. santo:57956-57957 for the round ending 21:40:50 — the 21:41 incident round — and santo:201693 for the round ending 22:47:21 — the 22:48 incident round). Michael (d:153225-153226) and Dave (a: tail before 218096) likewise ended the 22:47 round cleanly. No [steamnet] close, no ConsoleHost disconnect, no PEER_DOWN anomaly at either incident round-end on any logged machine. The freeze on the logged machines happened AFTER the last log line of a clean in-process mission end — i.e., in the relaunch/exit seam, NOT in Stop delivery. [T2]

But genuine console-link death happened that night — twice — and its signature matches the reported symptom exactly. Mid-round, all pods lost BOTH links to the host token (.1) with [steamnet] connection N closed (Timeout; remote problem. Rx age server 12-21s ...) followed by Disconnected from ConsoleHost at 169.254.77.1:1501: eleng f:22366-22372 (+ a second episode f:50374-50455) and the santo(e:35887-35922)/rajel(b:20340-20416) round (same round; rajel's clock is -3 h). After the loss the pod keeps simulating with no console and no fallback (finding 4): peers ghost out ([ghost] replicant 5:29 has received NO update records for 601 frames, f:~22434) and the mission never ends — each player eventually closed the window by hand ([marshal] window closed by the user -- exiting for real, e:36012, f:22510). The "Timeout; remote problem" reason means the HOST process went silent mid-round (crash/wedge, not graceful close) — the host is the machine with the marshal-thread pump race (finding 5). [T2; the race attribution is T3]

The frozen machines' actual signatures:

  • santo ~21:41: round ended cleanly 21:40:50 (e:57957); the menu child (pid 28112, e:57959-57966) booted fully, auto-joined the lobby within 1 s, and received GO normally at 21:43:10 (mission child pid 31404, relaunched generation, e:57967). The 2m20s hole between a clean end and the next GO, with a healthy-but-invisible menu, fits a wedged mission-parent window sitting frozen on top of the live menu (the foreground-handoff problem this code already documents, btl4console.cpp:262-271). [T2 logs; interpretation T3]
  • acaci 22:48: ZEUS's log goes SILENT at 22:12:38 — the post-round menu child (pid 29344, c:213897-213900) printed exactly 3 boot lines and never reached [steamnet] up, which for a relaunched child is unconditional ~1 s after boot (BT_STEAM_NET is inherited; boot-time install at btl4main.cpp:1306-1316). It hung at/before SteamAPI_Init (L4STEAMNET.cpp:337). acaci never played again (file mtime 23:18, nothing after 22:12:38) — acaci was NOT in the 22:37-22:47 round at all; yet acaci's SteamID (76561198064247614) still appears as token .2 in the 22:47 lobby map on santo/Michael (e:201702, d:153233) — a ghost lobby seat held by the wedged process. [T2; "hung at SteamAPI_Init" is T3 — the alternative, a user-kill inside the ~1 s boot window, is implausible given the same machine repeated it]
  • The same 3-line pre-SteamAPI_Init signature ends EVERY machine's log (Dave 22:47:22 — the exact incident round end, a:218096-218099; eleng 21:47:21; rajel 20:25:37; Michael 23:25:36; santo 23:25:37) and appears mid-night followed by ~1-2 min manual-restart gaps (Dave 20:31:52→20:32:40 a:31170; acaci 21:07:11→21:09:13 c:60457). The relaunch/exit seam wedges routinely; nobody calls SteamAPI_Shutdown anywhere (grep: SDK headers only), and every round-end exits via ExitProcess from BTFE_RelaunchSelfAndExit (btl4console.cpp:275) — on the HOST from the marshal thread while the main thread is mid-teardown (btl4console.cpp:583-584), matching Lynx's "hang on exit, hosting". [T2 signatures; ExitProcess/DLL-detach mechanism is T4]

4. Bottom line for #163

The StopMission delivery chain is a single unmonitored 10-minute-old connection per pod with at least four fully silent failure modes (send-result discarded; error mapped then discarded; reliable-buffer drop; dead-vs-empty indistinguishable in the drain loop) and no mesh-mode fallback on the pod — a real design hole, and the night's two host-death rounds show exactly what its failure looks like (orphaned forever-mission, frozen peers). However, for the two ticketed incidents the field logs exonerate delivery: every logged pod received Stop and ended its mission in-process; the freeze and the exit hang live in the round-end relaunch/exit seam (ExitProcess without Steam shutdown; menu children hanging pre-SteamAPI_Init; ghost lobby seats from wedged processes), on which the marshal's host-side thread-kill-at-clock-expiry piles extra risk for the host machine.

==============================================================================

#163 Field Forensics — night15 (2026-08-10, build 4.11.883)

0. Ground work: clock alignment and identity map (all T2 unless noted)

Wall clocks were aligned by matching identical [lobby] join ... map [...] strings and identical session-cadence across machines (per instructions, not by raw wall clock):

Machine User Steam ID Clock offset vs ET Evidence
MS-FIREFLY santo 76561198049064449 = ET own token, santo log line 15
SCREECH-PC Dave 76561198022091594 = ET sessions 21:30:18/21:40:54… match santo; hosted the 19:40 2-pod round (Dave 766) whose lobby shows .1=…022091594 (eleng line 9)
XIAOLONG Michael 76561198020775163 = ET sessions 21:43:10/21:53:46 match santo; hosted the 23:15 round (Michael 209577) whose map has .1=…020775163 (santo 255102)
DESKTOP-QR9VPJQ Lynx 76561198659597127 = ET matchlog B HDR (his machine) is exactly +1:00:00 vs ALIA's matchlog A of the same events
ALIA eleng 76561198147980449 ET − 1:00:00 glass sessions 20:30:17/20:43:10/…/21:36:48 = santo's 21:30:17/…/22:36:48 to the second
ZEUS acaci 76561198064247614 ET − 1:00:13 (≈ALIA−13s) ZEUS 21:36:35→21:47:11 pairs with ALIA 21:36:48→21:47:21
GAMERSLAB rajel 76561197970523304 ET − 3:00:00 rajel 18:40:50/18:43:10/18:53:46… = ALIA 20:40:50/20:43:10/20:53:47 exactly −2h (⇒ −3h vs ET, matching the "~3h behind" note)
(no log) ? 76561197976435895, 76561197995508393 — extra players, day logs not collected

CRITICAL CORRECTION TO THE TICKET'S PREMISE [T1]: matchlog_20260810_213656_20476.txt (ALIA, local 21:36) and matchlog_20260810_223651_12120.txt (DESKTOP-QR9VPJQ, local 22:36) are the SAME mission — the 22:36 ET round — recorded on two machines whose clocks differ by exactly 1h. Proof: identical MISSION run instant (A line 22: MISSION … w=21:37:21.041 st=4 run vs B line 22: w=22:37:20.047), identical terminal events (A line ~1319: DEATH victim=2:20 … pos=-27.5,0.0,144.2 + final SBMIRROR player=8:1 kills=4 deaths=2 wasKills=3 wasDeaths=2; B lines 1816-1818: same death at pos=-27.6,0.0,144.2, same SBMIRROR). Matchlogs are written per-machine (each pod keeps one), confirmed by dev-bench pairs in scratchpad/night15/mlbak/ (two matchlogs per bench round). There is no matchlog for the 21:41-incident round in the staged set. Lynx DID host the 22:36 round (host token .1=…659597127 in the round's map, ALIA line 172106) — and in fact hosted every round from ET 21:08 through 22:47, including BOTH incident rounds (santo maps at 36020, 46238, 57965, 68589, 101519, 124366, 156700 all show .1=…659597127).

1. The mission-end ladder as it appears in these logs

There are no 'EndingMission'/'StopMission'/state-name lines in this build's day logs. The observed healthy end shape (every clean round, every machine) is:

[mission] solo game clock expired (600s|1200s) -- ending the mission
[score] gauge read / ~RankAndScore: I think my score is N
[lamp] 0xNN <- 0x0            (×41, cockpit lamps off)
SVGA16::~SVGA16: pixel management statistics + 4 lines
[glasswin] destroy entry #1 windows=0
[glasswin] destroy entry #2 windows=0
[boot] RunMissions returned (mission loop exited).
[fe] mission over -- relaunching the menu        <=== LAST LINE the process ever logs
===== BT411 SESSION ... =====                    (the relaunched MENU process, new pid)

(e.g. santo 57904→57959). ~50-60 lines, wall time ≤3s (the menu header lands the same/next second as the computed 600s expiry). After [fe] mission over the process calls BTFE_RelaunchSelfAndExit("") → CreateProcessW + ExitProcess(0) (game/btl4main.cpp:1671-1673, game/glass/btl4console.cpp:157-275) — everything after the [fe] line is unlogged by construction.

2. Incident A — ~21:41 ET, santo (MS-FIREFLY) frozen

Round: ET 21:30→21:40:50, Lynx-hosted, 6 players (santo 46238 map: Lynx.1, eleng.2, acaci.3, santo.4, Dave.5, rajel.6), 600s clock. Michael not in it.

santo's end sequence is textbook-clean and on time [T1]:

  • 57887-57903: normal in-mission ticks (santo idle: speedDemand=0, alive — targeting active; the last nearby deaths are replicants 7:4 at 54956 and 6:4 at 57832, NOT santo).
  • 57904 [mission] solo game clock expired (600s) → 57906 ~RankAndScore: I think my score is 2686 → lamps → 57948 SVGA16::~SVGA16 → 57954-5 glasswin destroys → 57956 RunMissions returned → 57957 [fe] mission over -- relaunching the menu → 57959 new MENU session at 21:40:50 (same second as everyone else, below). The log never goes silent mid-ladder and never keeps ticking past the end. (No [glassperf] beats exist on this machine — the tag appears only in Dave's and Michael's logs.)
  • Recovery: the relaunched menu was alive immediately — 57963-57965 it joined the NEXT round's lobby (token .7); santo was playing again at 21:43:10 ET (57967), ≈2m20s after the freeze report. No manual cold restart visible.

Same-round healthy comparison (identical ladder shapes, same wall second):

machine EXPIRED line menu-header line menu local time (→ET)
santo (frozen) 57904 57959 21:40:50
eleng 73366 73424 20:40:50 (21:40:50)
acaci 11986 12045 20:40:37 (≈21:40:50)
Dave 75044 75126 21:40:54
rajel 44983 45041 18:40:50 (21:40:50)

Conclusion [inference, flagged]: the freeze left NO trace in santo's day log. Everything up to and including the menu relaunch executed on time; the only unlogged window is the old glass process's ExitProcess(0) path after [fe]. The "frozen view of how the game ended" is consistent with the OLD process hanging after spawning the menu — its last rendered frame stays on the fullscreen/plasma window on top while the new menu opens BEHIND (the build even logs [marshal] foreground handoff denied -- the next window may open behind for this Z-order hazard, e.g. santo 46239; emit site btl4console.cpp:267-271).

3. Incident B — ~22:48 ET, acaci (ZEUS) + Lynx (host) frozen

Round: ET 22:36:51→22:47:2x, Lynx-hosted, 8 players (map ALIA 172106: Lynx.1, acaci.2, …976435895.3, santo.4, eleng.5, Michael.6, Dave.7, rajel.8), GO at 22:37:20.0 (matchlog B line 22), 600s ⇒ nominal expiry 22:47:20 ET.

acaci (ZEUS) end sequence [T1] — clean, but with the night's most interesting receipt:

  • 159680 [wreck] replicant 2:20 entered wreck state … at (-27.6352,144.223) — Lynx's mech dying at 22:47:15.6, the same death both matchlogs end on.
  • 159890 [steamnet] connection 1 closed (closed) + 159892 Disconnected from GameMachineHost at 169.254.77.1:1502 — the HOST (Lynx) dropped ZEUS's game link BEFORE ZEUS's own clock expired (ZEUS's 600s clock ran ~3-10s later than the pack because its RunMission receipt timing; it was the last machine still in-mission).
  • 159894 EXPIRED (600s) → 159909/159911 second disconnect (169.254.77.5:1502 = eleng, tearing down in parallel) → 159970 RunMissions returned → 159971 [fe] → 159973 menu session 21:47:11 local (≈22:47:24 ET).
  • Recovery: menu joined the 22:52-round lobby immediately (159979); acaci played the 22:52 ET round (159981, glass 21:52:09 local) — recovered ≤5 min; quit for the night cleanly at 23:14:36 ET (lastrun_steam_acaci_ZEUS_8.txt line 29: [2026-08-10 22:14:36] clean exit: player quit from the menu; his day log's final 3-line menu stub at 213897 is the systemic buffered-tail loss described in §6, not a wedge).

Same-round healthy machines: Michael 153171→153227 (menu 22:47:20), santo 201637→201695 (22:47:21), rajel 192406→192463 (19:47:21=22:47:21), eleng 216583→216639 (21:47:21=22:47:21), Dave 218029→218096 (22:47:22). All within 0-4s of nominal expiry, identical ladders. eleng then quit for the night at 22:47:48 ET (lastrun line 47) — unaffected.

Lynx (host, no day log): his own matchlog (B) stops at 1818 SBMIRROR … w=22:47:15.805 — his own death (PLAYER_DEAD … player=2:1 deaths=10 at 22:47:15.642, B line ~1813), 5.6s before expiry, respawn never recorded — and contains no mission-end receipt. He reported "BTL4 hang on exit, hosting" + Task Manager; he is absent from both later lobbies (22:52 map santo 201701; 23:15 map santo 255102) — never returned. The pre-expiry link close seen on ZEUS ((closed), i.e. orderly close, not timeout) proves Lynx's process DID start its teardown (its clock expired first); the hang happened later in his exit path [inference].

4. Did StopMission ever arrive? (Task 4)

No — on any machine, at any round, all night [T1]:

  • grep StopMission over all six day logs (≈90MB): 0 hits.
  • Every mission end on every machine is the LOCAL fallback: [mission] solo game clock expired (Ns) -- ending the mission (14 on santo, 10 on eleng, 8 on acaci, etc.).
  • The marshal's only sends all night are launches: [marshal] RunMission #1 sent to N pod(s) / RunMission #2 sent -- mission running; clock Ns (Dave 766/899, 13401/13588, 16085/16289, 23146/23304, 32151/32319, 50684/50831; rajel 19145/19269; Michael 615/733, 209577/209723). No [marshal] StopMission line ever.
  • Consequence: rounds end as N independent local expiries skewed by RunMission-delivery jitter (0-13s spread observed), and the host tears its links down whenever ITS clock fires — peers see Disconnected from GameMachineHost … :1502 pre-expiry when their clock is late (ZEUS 159890 in incident B; also Dave 118701 & 141778, Michael 13749, eleng 60403 & 84435 — all Lynx-hosted rounds).
  • Corroborating shape from earlier the same night: when a HOST vanished mid-round (Dave closed his hosting window at 20:21:27 and 21:06:28 ET), peers did NOT transition — they sat in-mission until SteamNetworkingSockets timed out (Timeout; remote problem. Rx age server 11-21s: rajel 20340-20416, santo 35887-35922, eleng 22366-22440 & 50374-50455) and each user had to close the window (RunMissions returned + [marshal] window closed by the user). There is no host-loss or console-driven end path under Steam in the field [T1].

5. Matchlog tails vs bench baseline (Task 3)

  • Field matchlogs (both machines, incident-B round) end mid-combat with NO end receipt: A ends SBMIRROR … w=21:47:16.821, B ends SBMIRROR … w=22:47:15.805. Neither contains any st=6/stop/PEER_DOWN record; the only MISSION line each is st=4 run.
  • Dev-bench matchlogs (mlbak/, builds 874-885) normally end with PEER_DOWN t=… host=1 type=3 (the console/relay link closing at teardown) — e.g. matchlog_20260811_005558_2240.txt last line, matchlog_20260810_102334_9548.txt last line.
  • So both field matchlogs lost their teardown tail. Since ALIA demonstrably transitioned cleanly, the missing tail is NOT diagnostic of the hang; the likely mechanism is the ExitProcess(0) relaunch path discarding the matchlog's unflushed stdio buffer (btl4console.cpp:275) [T3].

6. Systemic log-tail artifact (don't misread it)

Every collected day log ends with a fresh MENU session header + exactly 3 [boot] lines and nothing more (Dave 22:47:22, eleng 21:47:21, rajel 20:25:37, Michael 23:25:36, acaci 22:12:38, santo 23:25:37). The lastrun files prove these menus ran fine and exited cleanly minutes later (ZEUS quit 22:14:36 local; ALIA quit 21:47:48 local; santo's final sessions rotated into steam_20260810.1.log — not collected; Dave's post-22:47 activity, including hosting the 22:52 ET round per map token .1=…022091594, is likewise missing from the collected file). The [boot] lines carry explicit std::flush (btl4main.cpp:1605-1619); the later [steamnet]/[lobby] lines evidently don't survive the quit path's buffer loss. Truncated final menu stub ≠ wedge.

7. Healthy baseline round (Task 1 control): ET 22:22:36 → 22:33:11, Lynx-hosted, 8 players

machine glass session (local) EXPIRED line menu line menu local (→ET)
santo 22:22:36 (124368) 156637 156694 22:33:11
Michael 22:22:36 (74169) 107656 107712 22:33:10
Dave 22:22:37 (141879) 173829 173890 22:33:11
eleng 21:22:36 (140157) 172042 172100 21:33:11 (22:33:11)
acaci 21:22:23 (83362) 115098 115152 21:32:57 (≈22:33:10)
rajel 19:22:36 (113680) 146649 146702 19:33:11 (22:33:11)

Identical ~55-line ladders, all six menus stamped within 1s. The incident-round ladders of the frozen machines are line-for-line indistinguishable from this baseline.

8. Synthesis (clearly marked inference where noted)

  1. [T1] The in-log mission-end ladder COMPLETED normally and on time on every frozen machine; the freeze lives entirely in the unlogged post-[fe] exit path (BTFE_RelaunchSelfAndExit → ExitProcess(0), btl4console.cpp:275).
  2. [Inference] "Frozen view of the ended mission" = the old glass process hung inside ExitProcess (classic DLL-detach/terminated-thread deadlock territory: SNS threads, D3D, the plasma-window thread) with its last frame still on screen, while the already-spawned menu opened BEHIND it (AllowSetForegroundWindow hazard logged at btl4console.cpp:269). Task Manager kill of the OLD pid clears it — exactly Lynx's report.
  3. [T1] Both incidents happened in Lynx-hosted rounds; Lynx's own machine hung at exit in incident B while hosting (7 console links + relay + marshal to unwind). In incident B, Lynx's teardown began BEFORE the last peer's expiry (orderly (closed) on ZEUS 159890) and then hung.
  4. [T1] StopMission is not part of this build's field behavior — every end is the local solo clock; the marshal only launches. Host death mid-round strands every peer in-mission until manual window close (three occurrences 20:08-21:07 ET).
  5. [T2] Lynx died in-game 5.6s before the clock and his respawn never appears before the end; acaci and santo were alive at their freezes — a death-at-end race is NOT the common factor across the three frozen instances.
  6. Recovery: santo ≈2m20s (auto-relaunch worked; next round 21:43:10 ET); acaci ≈5 min (next round 22:52 ET, quit cleanly 23:14:36 ET); Lynx: killed via Task Manager, never rejoined.

==============================================================================

Ticket #163 / #156 — TASK 3: The binary's end-of-mission truth

Scope: the 1995 BTL4OPT.EXE end ladder, read from reference/decomp/all/part_*.c + the T0 WinTesla MUNGA source the port compiles against, with constants byte-read from the image (scratchpad/rdva.py) and one raw disasm (tools/disas2.py). Everything is [T1] (decompiled + cross-checked against T0/raw bytes) unless tagged.

0. Headline answer

The 1995 end sequence is EVENT → TIME → EVENT, and the only timed leg runs inside the mission player's own simulation tick. Console StopMission (event) puts the app in EndingMission (state 6) and dispatches MissionEnding to the mission player; the player then counts a 3.0-second fadeTimeRemaining down in player-Performance time (ManageApplicationStatus @0x42df8c); at ≤0 it self-dispatches a second StopMissionMessage (event) which, arriving at state 6, calls Application::Stop() @0x44e6c0 → executeFrames=0 → the mission loop unwinds. Nothing in the ladder waits on the visual fade — the fog fade renderable (@0x45447c) is a passive video watcher. The ladder CAN stall in exactly two binary-grounded ways: (a) StopMission #1 never arrives — the mission runs forever (1995's console owned the clock; the pod never stops itself — context/multiplayer.md:639-641); (b) the mission player's Performance never executes while in MissionEndingState — then fadeTimeRemaining never decrements, StopMission #2 is never dispatched, and the app parks in EndingMission FOREVER. Once the countdown completes, StopMission #2 is a direct synchronous application->Dispatch (PLAYER.cpp:454-455; @0x42df8c line 740) — it cannot be lost in transit.

1. The Application state machine's EndingMission arm (binary addresses)

State enum (T0 engine/MUNGA/APP.h:188-202): Initializing=0, WaitingForEgg=1, LoadingMission=2, WaitingForLaunch=3, LaunchingMission=4, RunningMission=5, EndingMission=6, StoppingMission=7, Suspending=8, Resuming=9, Aborting=10, CreatingMission=11. App state read as *(app+0x88); the StateIndicator sits at app+0x74 with current state at indicator+0x14, and SetState = FUN_0041bbd8 @0x41bbd8 (part_002.c:5510) — which also pings three watcher chains (indicator+0x18/+0x2c/+0x40, vtbl+0x14 notify) — this is how the fade renderable learns of state flips.

Binary Application TU (part_007.c:3090-3810):

  • FUN_0044eeb4 @0x44eeb4 = Application::StopMissionMessageHandler (part_007.c:3698-3737): state 7 → ignore; state 6 or 10 → Stop() (@0x44e6c0) — this is the SECOND arrival; default → SetState(6), and if missionPlayer(app+0x14) non-null: networkManager->Mode(0=ReliableMode) (vtbl+0x30 on app+0x20, part_007.c:3721 == T0 APP.cpp:1762) then dispatch Player::MissionEndingMessage {size 0x1c, id 0x19} to the player; if NO mission player → Stop() directly. Matches T0 APP.cpp:1705-1778 exactly.
  • FUN_0044ef4c @0x44ef4c = AbortMissionMessageHandler — same shape, SetState(10), no ReliableMode switch.
  • FUN_0044e6c0 @0x44e6c0 = Application::Stop(): executeFrames(app+0x70)=0; flush DEBUG stream; SetState(7). (T0 APP.cpp:797-821.)
  • FUN_0044e488 @0x44e488 = Application::ExecuteForeground (part_007.c:3138-3176): returns 0 immediately when executeFrames==0; at state 5 recomputes secondsRemainingInGame(app+0x58) = mission->length(+0xe0) − (Now−gameStarted(+0x5c))/tps; final return is executeFrames && !Exit_Code where Exit_Code = DAT_004efc98. There is NO pod-side action when the clock hits 0 — app+0x58 is display-only; the stop was the console's job.
  • FUN_0044e6e8 @0x44e6e8 = base Application::Shutdown (part_007.c:3239-3278): gauge/video/audio renderer Shutdown+UnlinkFromEntity (+0x4c/+0x48/+0x44), delete viewpointEntity (+0x6c), interestManager/hostManager/networkManager shutdown, delete currentMission(+0xC8), then executeFrames=1; SetState(1=WaitingForEgg); return 0.
  • Other rungs for contrast: CheckLoad @0x44ebec (2→3, and if no console host (FUN_00429078(app+0x2c)) self-launches RunMission — the console-less path), RunMission @0x44ecdc (3→4 + dispatch MissionStarting id 0x18 to player; 4→5 + gameStarted=Now(); else Fail @APP.CPP:0x609 — the port's launch/load-race crash site), Suspend @0x44ed88 (5→8 + dispatch id 0x19 — suspend ALSO rides the player MissionEnding fade), Resume @0x44ee10, KeyCommand @0x44efd8 ('&'=0x26 → Exit_Code=1 + Stop()).

There is NO app-side timer in the end arm. The only end-sequence time constant anywhere is the player's 3.0 s fade (plus the fade renderable's cosmetic 1.0 s, §3). No mission-review display period exists at the app level; the ranking window is a separate per-player display driver (§below).

Ranking-window behavior at end — FUN_0042eb38 @0x42eb38 (part_004.c:1387-1426, called from PlayerSimulation/CameraShipSimulation): at states 6/7 the window flag (player+0x224) is forced 0 (hidden); while secondsRemainingInGame > 30.0 (_DAT_0042ec2c=30.0, byte-read) it runs a periodic show/hide machine (+0x228/+0x22c/+0x230/+0x234); at ≤30 s it pins the flag to 1 (solid on). So the authentic end was: standings solid for the final 30 s, then StopMission → ranking hidden + 1 s fade to black. (The camera-seat DIRECTOR.cpp:107 additionally shows its ranking during EndingMission — different display, camera seat only.)

Crucial executability fact [T0]: Entity::Execute (ENTITY.cpp:556-558, quoted in context/reconstruction-gotchas.md §29:920-934) calls PerformAndWatch only in RunningMission || EndingMission || IsPreRunnable() — so the world (and the player's fade countdown) KEEPS TICKING during state 6, and stops at state 7. A screen showing a frozen last frame therefore means the app got PAST state 6 (Stop() ran, executeFrames=0, renderers no longer execute) but the process never finished Shutdown/relaunch — whereas a player stuck in state 6 would see a live, still-simulating world that never returns to menu. (Relevance flag for #163 triage, [T4 inference]: post-RunMissions in the modern port comes the matchlog-upload dial + relaunch (context/multiplayer.md:110, :541-545) — a blocking network call there matches "frozen final frame + hang on exit".)

2. The Player handlers @0x4bfc20 / @0x4bfbe8

Base handlers (part_004.c:409-430):

  • FUN_0042d9c0 @0x42d9c0 = base Player::MissionStartingMessageHandler: fadeTimeRemaining(+0x1F4) = 3.0f; SetSimulationState(3=MissionStartingState).
  • FUN_0042d9e0 @0x42d9e0 = base Player::MissionEndingMessageHandler: fadeTimeRemaining = 3.0f; ForceUpdate (+0x18 |= 1) — this ships the final player update record (score) under the just-switched reliable mode; SetSimulationState(4=MissionEndingState). Neither touches the Performance pointer or simulationFlags.

BT overrides (part_013.c:18556-18588; raw disasm of @0x4bfc20 via tools/disas2.py this session):

  • FUN_004bfbe8 @0x4bfbe8 = BTPlayer::MissionStarting: base @0x42d9c0, then if (app->state==4 /*LaunchingMission*/ && !(this+0x29 & 0x40 /*NonScoringPlayerBit*/)) currentScore(+0x1c8) = 1000.0f — the manual's "+1000 starting the game" seed (port: game/reconstructed/btplayer.cpp:788-806).
  • FUN_004bfc20 @0x4bfc20 = BTPlayer::MissionEnding — full body:
    1. base @0x42d9e0 (fade 3.0 + ForceUpdate + state 4);
    2. this+0x288 −= Round(15.0f × DAT_0052140c + 0.5) — recovered from raw bytes @0x4bfc3b-0x4bfc52 (fld 15.0f @0x4bfca4; fmul [0x52140c]; fadd 0.5f @0x4bfca8; call __ftol@0x4dcd94; sub [ebx+0x288],eax): it rewinds the 15-second console score-total clock by exactly one period, so the next PlayerSimulation tick immediately pushes the FINAL total score to the console (see §5). The old export dropped the x87 operand (FUN_004dcd94() rendered argless — the §19 gotcha); now pinned.
    3. if (app->state == 8 /*SuspendingMission*/ && playerVehicle(+0x1fc) IsDerivedFrom Mech (tag 0x50bdb4)) → FUN_0049fb74(mech, &DAT_00524b38, 1) (= Mech::Reset, per the KB respawn chain — context/decomp-reference.md "Respawn SIM path"; the argument blob DAT_00524b38 semantics unverified [T4]) + ForceUpdate. So the operator SUSPEND path resets the vehicle; the normal state-6 path does NOT.
    • No score finalization, no fade rendering, no waiting happens in this handler — it is fire-and-forget bookkeeping; the actual fade/countdown/exit all run elsewhere.

The countdown — FUN_0042df8c @0x42df8c = Player::ManageApplicationStatus (part_004.c:706-752; == T0 PLAYER.cpp:408-476): state 3: fade −= dt; at ≤0 (_DAT_0042e0dc=0.0) → state 2 (VehicleTranslocated) + dispatch RunMission (app state 4) / Resume (state 9). State 4 (MissionEnding): fade −= dt; at ≤0 → app state 6 → dispatch StopMissionMessage(exit 0) (ctor @0x44f5d0); state 8 → SuspendMission; state 10 → AbortMission. Called every tick from Player::PlayerSimulation @0x42e100 and Player::CameraShipSimulation @0x42e0e0 (part_004.c:756-781) — both flavors complete the ladder. (KB nit: btplayer.cpp:1329 cites base PlayerSimulation as FUN_0042e168; @0x42e168 is only the vehicle-position-copy tail — the full base is @0x42e100.)

When can the countdown not run? Simulation::DoNothingOnce = NeverExecute() = simulationFlags |= DontExecuteFlag (SIMULATE.cpp:484-488, SIMULATE.h:167-206). In the 1995 binary the master player is parked on DoNothingOnce only pre-first-spawn (Player::HuntForDropZone @0x42ddcc sets the pmf PTR_FUN_004e67fc at part_004.c:669 once the drop zone is found, until DropZoneReply re-arms). The BT death cycle does NOT park it: FUN_004c05c4 @0x4c05c4 = BTPlayer::VehicleDead (part_013.c:19064-19156) contains no SetPerformance/simulationFlags writes — a mid-mission-dead player still ticks PlayerSimulation, so a StopMission arriving during a death window still completes the fade in 1995. Its first line is the state-4 gate (this+0x40==4 → return): once MissionEnding, all death traffic is swallowed. Also in the -1 arm: deaths dual-increment (+0x280), console MechKilled id 8 via ctor @0x4c18cc gated suppressConsole(+0x258)==0, objectiveMech(+0x284)=killer's vehicle, lives check role(+0x208)->returnFromDeath(+0x28) < 1 → post Mech id 0x18 (ClearBurningState) to the wreck at +10 s and NO respawn re-post; else re-post VehicleDead to self at +5.0 s (@0x4c0830=5.0f, byte-read); then the straight-line death cost gated advancedDamageOn(+0x264) via direct call to base ScoreMessageHandler @0x42da20.

3. THE FADE (ticket #156)

Mechanism: a fog-color/fog-range animation on the main view — POVStartEndRenderable (T0 engine/MUNGA_L4/L4VIDRND.cpp:2126-2394, binary ctor FUN_00454394 @0x454394, Execute FUN_0045447c @0x45447c, part_007.c:8468-8606). Not a palette fade, not a gauge overlay, not the warp machinery.

  • Built for the LOCAL (master) player only by the video builder FUN_004d0774 @0x4d0774 (part_014.c:10673-10693): classID 0xBDA (BTPlayer): replicant ((entity+0x28 & 0xc)==4) → translocation-warp renderable @0x458d2c; master → alloc 0x50 → FUN_00454394(entity, Watcher=2, mainView, mainZone, deathZone, trigger, fogRGB, fogNear/Far, 3, 4). The trigger = the player's SimulationState attribute (FUN_0041bf44(entity,1)), i.e. the same StateIndicator the MissionStarting/Ending handlers write; start state 3 = MissionStartingState, end state 4 = MissionEndingState.
  • State machine (@0x45447c; myState at +0x1c, timer +0x3c, trigger state read at trigger+0x14): 0 WaitForStart → on trigger==3: fog snapped to WHITE (1,1,1) with near/far 0.01/0.05 (screen fully fogged = white flash), AddDynamicRenderable; 1 FlashScreen: hold 0.1 s (@0x454740); 2 FadeIn: lerp fog color/ranges back over 1.0 s (@0x454744); 3 MissionRunning: wait for trigger==4 → 4 FadeOut: fog color × percent-left → black, ranges collapse to 0.01/0.05 (@0x45474c/0x454750), over 1.0 s in the binary — note the WinTesla source's FADE_OUT_TIME (0.5f) (L4VIDRND.cpp:2150) is a later edit; the shipped 1995 constant is 1.0 (the same @0x454744 is reused as the fade-out duration at part_007.c:8584); at expiry → back to state 0 + RemoveDynamicRenderable.
  • What advances it: the renderer's frame clock (GetCurrentFrameTime), NOT the player sim — once state 4 is set it completes even if the player sim is dormant. What would happen if it never completed: nothing. The ladder never reads the renderable; ManageApplicationStatus is the sole gatekeeper. The fade-out (1.0 s) simply fits inside the 3.0 s window; the pod then sat on a black screen ~2 s before teardown.
  • Why the port has no fade (#156 root cause): the builder call exists (game/reconstructed/btl4vid.cpp:311-315, states 3/4 passed), but BTPOVStartEndRenderable's ctor is an empty stub — game/reconstructed/btstubs.cpp:339-342 — it never registers with the video renderer or the state dial. The white mission-start flash is missing for the same reason. All the plumbing it needs (SetState watcher ping @0x41bbd8, fog style/limits on the D3D renderer) already exists in the port.

4. RunMissions exit + what the 1995 pod did next

  • Exit condition (T0 engine/MUNGA/APPMGR.cpp:39-161): RunMissions loops foreground/packet-route/background passes; when application->ExecuteForeground(...) returns False (binary: executeFrames==0 || Exit_Code(DAT_004efc98) — @0x44e488) it calls application->Shutdown(n); Shutdown returning False removes the app from runningApplications, and with no apps left RunMissions Terminates them and RETURNS.
  • 1995: the pod did NOT exit — the L4 (pod) application overrides Shutdown: FUN_0047c560 @0x47c560 (part_010.c:2805-2820) = base Shutdown @0x44e6e8 (which re-arms executeFrames=1 + SetState(WaitingForEgg)) then return (Exit_Code==0 && app+0xd4==0) ? 1 : 0 — i.e. the app STAYS in the manager and the pod parks in WaitingForEgg for the console's next egg (the between-rounds attract wait). RunMissions only returned on the '&' keystroke (Exit_Code, @0x44efd8) or the app+0xd4 shutdown latch (set by the two-phase handler @0x47c2c4 — first receipt latches +0xd4 and re-posts, second receipt tears down; a console-initiated pod shutdown [T3 on its exact message id]).
  • The modern build diverges deliberately: T0 APP.cpp:922-926 replaced the original return !Exit_Code with a hard return False (#if 0 preserved in-source), and engine/MUNGA_L4/L4APP.cpp has no Shutdown override — so every mission end exits RunMissions, then btl4main uploads the matchlog and BT_FE_LOOP relaunches the menu (context/multiplayer.md:110,541-545).

5. The end-of-mission pod↔console handshake

All pod→console, fire-and-forget over the console link (clientID 5 = ConsoleClientID); the ladder never waits for any console reply. Names from the surviving original game/original/BT/BTCNSL.CPP (cnslmsgs.cpp) + T0 engine/MUNGA/CONSOLE.h:11-18:

When Message {size, id} Binary ctor / site
Every 10 s AND immediately once app state == 6 score DELTA flush: self-dispatch ScoreUpdate {0x20, 0x1a} → the +0x278 delta cell is sent then zeroed FUN_004c083c @0x4c083c (part_013.c:19160-19192): 10.0(@0x4c08fc) <= (this+0x10 − this+0x28c)/tps || app+0x88==6; port analog btplayer.cpp:1361-1427
Every 15 s while state 5 or 6; forced immediately at MissionEnding by the @0x4bfc20 clock rewind ConsolePlayerMechScoreUpdate (running TOTAL, from +0x1c8) {0x14, id 9} ctor @0x4c18f4; send site part_013.c:19269-19279 (_DAT_004c0bc0=15.0, clock at +0x288)
Once, when fadeTimeRemaining ≤ 1.5 in MissionEndingState (i.e. halfway through the 3 s fade, while the link is guaranteed still up), one-shot latch +0x254 state 6: ConsoleApplicationEndMissionMessage {0x14, id 7} {host, Round(score)} (added 10/05/95 per BTCNSL.CPP:10); otherwise (abort/suspend): ConsoleApplicationAbortMissionMessage {0x10, id 0xb} BTPlayer::PlayerSimulation FUN_004c0904 @0x4c0904, block part_013.c:19251-19268, Tell string "Sending EndMission..." @0x5132d7; ctors @0x42998c / @0x4299b4
On death (any time) ConsolePlayerMechKilled {0x14, id 8} {victim, killer} ctor @0x4c18cc, sent from VehicleDead @0x4c05c4 (part_013.c:19110-19114), suppressed after eject via +0x258
On eject DeathWithoutHonor (id 5) FUN_004c198c (KB: decomp-reference EJECT cluster)
(team variants) TeamScoreUpdate {0x14, 0xd} @0x4c191c; TeamEndMission {0x1c, 0xe} @0x4299d8

Plus two non-console end actions: networkManager->Mode(ReliableMode) at StopMission #1 (@0x44eeb4/T0 APP.cpp:1762) and the base MissionEnding ForceUpdate shipping the final player record to peers. Also note part_013.c:19280-19282: while state 5 or 6, the player record is re-dirtied (+0x18 |= 1) whenever none has shipped for 1.0 s (_DAT_004c0bc4) — a 1995 1 Hz player-record keepalive that keeps running through the ending window.

6. Stall matrix (the ticket's question, binary-grounded)

Leg Driven by Can it stall?
StopMission #1 → state 6 console event over the console link YES — if it never arrives the mission runs forever (1995 design; modern relay adds the +15 s RelayGameDown fallback, multiplayer.md:519-521). World stays LIVE (not frozen).
state 6 → StopMission #2 3.0 s of MISSION-PLAYER SIM TIME (@0x42df8c) YES — iff the mission player's Performance stops executing while in MissionEndingState (DontExecuteFlag latched, or its Performance never calls ManageApplicationStatus). In 1995 the only such window is pre-first-spawn (HuntForDropZone→DoNothingOnce, part_004.c:669). Result: app parks in state 6 forever — world keeps ticking (ENTITY.cpp:556-558), ranking window hidden (@0x42eb38), never returns to menu.
StopMission #2 → Stop() direct synchronous Dispatch NO (cannot be lost).
Stop() → RunMissions return next foreground pass: ExecuteForeground→0, then Shutdown (synchronous renderer/network/mission teardown @0x44e6e8) Only if a Shutdown callee BLOCKS (e.g. a network shutdown on a dead transport). From Stop() onward the screen is a FROZEN LAST FRAME (renderers no longer execute) — the shape #163 describes.
visual fade renderer frame clock, passive watcher irrelevant — nothing waits on it; in the port it's a stub anyway (btstubs.cpp:339-342).
Reference in New Issue View Git Blame Copy Permalink