{
  "summary": "Windows WMI escape itself succeeded; the 6.288 s failure is exclusively the test's `daemon stop` exit-success-as-reachability oracle. Strongest constrained prediction is ambient endpoint-identity/rig-class refusal, not exhausted load deadlines or the candidate input changes. Parent reports explicit OWL_SESSION_ID stop refusals in sibling failures; this cell discards stderr, so its exact refusal remains inference. Linux does not compile or exercise this Windows cell.",
  "files": [
    {
      "path": ".worktrees/asm-304-v3/crates/spt/tests/job_escape_e2e.rs",
      "description": "Windows-only job/WMI scenario, readiness/survival/stop/reap bounds; Unix names a separate empty documented guard."
    },
    {
      "path": ".worktrees/asm-304-v3/crates/spt-daemon/src/daemon.rs",
      "description": "Service-aware start/stop, elevation pre-rung, WMI-first launch ladder, explicit SPT_* forwarding and detached console handling."
    },
    {
      "path": ".worktrees/asm-304-v3/crates/spt/src/cli.rs",
      "description": "Stop exit status includes endpoint and live-session policy refusal, not merely IPC reachability; ambient identity/ancestry gate."
    },
    {
      "path": ".worktrees/asm-304-v3/crates/spt-daemon/src/seedmap.rs",
      "description": "Stop dial and post-ack barrier bounds; stop is a stronger protocol operation than ping."
    },
    {
      "path": ".spt/preserved/304-handoff/consumer-windows/phase-b.log",
      "description": "Named failure at lines 706-727: WMI marker, lethal control, daemon survival, successful authenticated cleanup, reachable=false only."
    },
    {
      "path": "docs/KNOWN-HAZARDS.md",
      "description": "7.10 at line 1101 documents WMI/env/job-neutral contract; line 347 warns conhost probe/spawn-flag inference can be wrong."
    },
    {
      "path": "docs/INFRA-REGISTER.md",
      "description": "IR-63 lines 3599-3633 describes pool co-residence/random victims and job-escaped leaks; lines 4094-4098 document measured WMI env forwarding, explicitly not Unix/schtasks."
    }
  ],
  "architecture": "Real Windows CLI start → optional de-elevation → forced WMI creator → cmd environment wrapper → detached broker → supervised brain; kill job → identity-based survival probe → service/policy-aware CLI stop → authenticated tree reap. No rc attach or user-input transaction is performed.",
  "report": "OBSERVED FAILURE: permitted phase-b.log:706 reports FAIL 6.288s; :720-724 records pinned brain 57340, broker 46220 authenticated and killed with its descendants, then rung_marker=\"WMI\", control_alive_before=true, daemon_alive_before=true, control_died=true, daemon_survived=true, reachable=false, tree_reaped=true. Panic :726-727 points to test :608, the reachability assertion. Thus neither launch, actual job escape, nor cleanup is the observed failing claim. Stop output/status detail is unavailable: test :502-516 discards stdout/stderr and reduces `cmd.output()` to successful exit. The 20s outer stop receive deadline did not expire in a whole test lasting 6.288s; 40s readiness and 25s cleanup ceilings likewise were not exhausted.\n\nPLATFORM: test :31-32 gates common helpers to Windows; :38-44 defines a DIFFERENT Unix test `daemon_survives_terminal_close_via_session_detachment` with an empty body that documents setsid/lifecycle coverage elsewhere; :46-47 cfg(windows) excludes the entire win module. A Linux job or same-binary PASS is not Windows-cell evidence. Main found no matching Linux PASS line, fully consistent with source exclusion. Unix production spawn is its own setsid branch (daemon.rs:920 onward), not WMI.\n\nEXECUTED PATH AND BOUNDS: test :118-133 creates KILL_ON_JOB_CLOSE job; :138-147 treats ACCESS_DENIED assignment as SKIP; :382-431 starts/assigns ping control and `spt daemon start` with only SPT_HOME + FORCE_LAUNCH_RUNG=wmi + marker overrides. Assignment follows spawn, not suspended creation. This observed run did not skip. :433-449 polls brain.ready up to 40s every 100ms, then unbounded start.wait(); :473-499 pins brain identity, TerminateJobObject, then 30 x 50ms probes (nominal 1.5s) using control.try_wait rather than stale-handle OpenProcess liveness. :502-516 allows 20s for CLI stop, but collapses policy/error/exit outcomes. :540-575 kills broker tree with authentication, then pinned brain, and polls identity disappearance for 25s every 100ms. Unbounded subprocess waits mean these are local bounds, not a strict total test deadline.\n\nLAUNCH/ENV: daemon.rs:781-810 service-aware start clears inhibit, spawns and waits; :873-918 first tries de-elevation, which may precede/avoid the ladder, but the observed WMI marker proves ladder was reached. :1939-1964 defines first-success WMI → schtasks → breakaway → in-job; :1976-1986 forcing WMI supplies FOUR WMI attempts, not fallback coverage. :1995-2000 writes rung marker. :2118-2133 forwards all SPT_* values except those containing a double quote via `cmd.exe /c set \"K=V\" & ... start \"\" /b`; this is not wholesale inherited caller environment (OWL_* is not forwarded by this wrapper). :2150-2155 uses Win32_ProcessStartup DETACHED_PROCESS=8 + SW_HIDE=0. :2171-2217 invokes absolute PowerShell, UTF-16LE base64 EncodedCommand, no-window caller, requires ReturnValue=0 and nonzero pid; `.output()` has no local timeout. `--detached` broker performs FreeConsole and nulls invalid handles (daemon.rs:2012-2050; cli.rs:7957-7974), then installs persistent stderr sink.\n\nCONHOST IS A SIBLING, NOT THIS CELL'S ORACLE: helper test :152-171 runs absolute PowerShell Get-CimInstance parent-id query, with no subprocess timeout and maps query/parse errors to zero. Sibling :208-240 has 40s brain-ready + 10s broker-pid polls; :279-293 nominal 5s conhost polling every 200ms (queries can overrun), followed by a final count. Target cold-start cell never calls this helper. Adjacent phase-b.log:730 reports sibling PASS 4.117s; that cannot replace target stop evidence or validate every probe failure branch.\n\nAPI MISMATCH / MOST LIKELY ALTERNATIVE: cli.rs:8120-8135 first obtains live sessions via ping + brain connection. :8140-8156 refuses stop for endpoint identity/ancestry BEFORE attempting stop, independently of IPC health; :8168-8187 returns zero for NotRunning as well as successful service/IPC stop. Therefore `status.success()` neither exclusively demonstrates a responded seed stop nor interprets nonzero as unreachable. Identity resolver :24470-24476 accepts any nonempty OWL_SESSION_ID / SPT_AGENT_ID / SPT_ENDPOINT_ID, else hosted ancestry. Target stop child changes only SPT_HOME, preserving ambient caller markers. Shared run.ps1:19-27 sets profile/ephemeral ports/build jobs and directly invokes cargo nextest without identity scrub. Main reports sibling preserved failures explicitly `DAEMON_STOP_REFUSED` because OWL_SESSION_ID is set (Phase B:103-104 and Phase A:74-75; parent-provided evidence, not reread here). [INFERENCE] The identical ambient identity likely makes this target's fast stop fail while the healthy daemon survives. Exact target stop message is irretrievably discarded by the test, so this is a supported prediction, not proved RCA.\n\nOTHER STOP BOUNDS: daemon.rs:819-832 service-aware stop can return NotRunning, service stop, or request_stop. seedmap.rs:263-268 defines 5s post-ack barrier, 10ms polling, 2s bounded probes; :339-379 bounds initial dial to 2s and performs stop write/read and barrier. Barrier may overrun its nominal ceiling by an in-flight probe; plain ping and stop frame reads are not equivalent to fully deadline-bounded end-to-end CLI execution. Fast real IPC failure remains possible; elapsed alone cannot select it versus policy refusal.\n\nBATCH OVERLAP: No `rc` is launched (test :23-25 explicitly delegates full rc reattach elsewhere), so rc InputPaths/input_path_worker is not exercised. No API state/USER_INPUT report, user-input event, controller attachment, or delivery injection appears in the target; nowsignal report_user_input and broker InputRecord/deliverybytes physical writer do not receive target-driven input (broker.rs:4267-4271 creates per-session writers; :4342-4349 writes records; :6382-6386 handles USER_INPUT_REPORT). The broker and brain processes do start, but that is not evidence these input branches execute. PRESENCE_PROBE_CEILING belongs to unlisted-row querying (cli.rs:7002-7035), absent from daemon start/stop. #49/#267 broker controller-close log is gated on real occupied-controller transition plus diag_trace_enabled (broker.rs:2756-2766), not performed by this test. Brain promotion not-drained log additionally requires a not-drained ready trial and diag_trace_enabled (brainproc.rs:875-878, :911-918); production brain startup exists but does not establish that conditional diagnostic ran. xtask seed scrub is git subprocess `.env_remove(\"SPT_DEBUG_RELEASE_SEED\")` at crates/xtask/src/main.rs:2765-2769; this runner invokes cargo nextest directly, so scrub is not executed and would not clear OWL_SESSION_ID anyway.\n\nHISTORY / CONSTRAINED PREDICTION: KH 7.10 (docs/KNOWN-HAZARDS.md:1101) precisely matches WMI job-neutral forwarding/ladder. IR-63 (docs/INFRA-REGISTER.md:3599-3618) supports a general co-resident-load/rotating-victim prior, but its signatures are leak locks and differing sweep victims, NOT evidence that this specific fast policy-or-IPC exit is starvation. Its :3620-3633 orphan headless conhost/cwd face is also not this successful-reap failure. IR entry :4094-4098 separately measures SPT_HOME, SPT_DOCS_PORT and rig-flag forwarding through WMI and explicitly disclaims schtasks/Unix; corroborates environment contract, not candidate validation. No exact cold_start/job_escape row was found in docs/FLAKE-LEDGER.md by named and broader WMI/cold_start/kitsubito searches. Prefer rig/environment-class (specifically inherited caller policy identity) over candidate-input-class, NOT a blanket proven load flake. REFUTATION: a clean-identity/ancestry context with captured stop exit+stderr still reproducing the same failure would weaken the identity account; demonstrated real stop IPC failure linked to changed production path would support candidate-class. No reruns/validation performed, per strict read-only constraint."
}