---
name: b6-liverepro-hunt
description: B6/F-e force-native-decay live-repro (2026-07-03) — NOT reproduced across 5-cell matrix; reproduced recoverable pump-stall (F-a) + transient RC_FAIL; parked to counter-43 RC window as int
metadata: 
  node_type: memory
  type: project
  originSessionId: 6381498d-5efd-4af4-835f-b4ef61197264
---

**2026-07-03 (perri): doyle-directed B6/F-e live-repro of the force-native decay** (incident: force-native worked 02:29, refused 02:52 on detached daemon, translate pid ALIVE). spt-core 0.22.0, adapter 0.12.0, node HFENDULEAM. Evidence archived: `scratchpad/b6-evidence-archive/` (stderr + 4 probe tables + README matrix).

**VERDICT: force-native decay NOT reproduced.** Full matrix, all cells → force-native DELIVERS:
- dormancy-pure (foreground, 27min) 10/10; psyche-thrash (foreground, ~1600 boots/10min, F-h faithful) 12/12 no floor poison; detached×time (27min) 18/18; sustained rc-churn (detached, 30cyc) 30/30. env-static (detached CreateEnvironmentBlock) delivers.

**Key logic that refocused the hunt:** incident daemon was detached-env for its ENTIRE life yet delivery worked early / failed late → static env-root EXONERATED as trigger; the missing cell is TEMPORAL (detached × time) — but that came back clean too.

**Reproduced (banked, NOT the target bug):**
- **Recoverable pump stall (F-a):** sustained rc --take/--view churn stalls peer pump within 1 cycle; SELF-RECOVERS (~3.5min, no restart). Transient, not hard death. **force-native survived the stall → pump ≠ inject-path, independent casualties.**
- **RC_FAIL:brain-IPC-read-deadline:** one-shot transient on rc --take vs a stale-controller record (old-binary rc, no B1/B4); did NOT recur under 30-cycle churn.

**The one undriveable cell → int #4:** a REAL agent-fired /clear boundary (SessionStart→checkpoint_fire hook self-send, itself --force-native) on a churn-stalled pump = the incident's actual pre-decay op. Old-binary rc can't drive /clear (injected "/clear" read as text; --take → RC_FAIL). **Folded into counter-43 RC window as int #4** (B1/B4-fixed rc drives boundary cleanly), alongside B2 pair + C3 give-up.

**INT #4 SPEC SHARPENED (doyle+todlando, 2026-07-03) — bank, run in RC window:** root candidate = a CLI-side **BOUNDARY-TRANSITION RACE**. The SessionStart hook's checkpoint force-native evaluates `is_spt_hosted_no_relay`'s gate — `ready-file exists && status==ONLINE && controllable==Some(true)` — at the exact instant a real /clear rotates the perch. Broker-side binding EXONERATED (endpoint-keyed, clear-transparent). Int #4 adds a CAPTURE LEG: at the hook-fire instant during the real /clear, record (a) target perch info.json status + ready-file existence + controllable, and (b) `detect_self_id`'s resolved target id. **Transient not-online / not-ready / unset-controllable / wrong-id at that instant = root CONFIRMED** (re-stamp vs hook-timing race). All-healthy + still-refused = back to doyle with that data. Rest of int-4 plan unchanged.

**Harness lessons (reusable):**
- Foreground daemon for stderr capture is SELF-DEFEATING with a normal-env client if envs diverge — but the real killer was: **any `spt` CLI call auto-starts a detached daemon when none runs** ("any spt command auto-starts it"), racing a foreground launch. Once a daemon IS up, CLI calls attach to it (no race). So: launch daemon, then never `spt daemon stop/start` mid-run.
- **F-026 sharpening:** trust pre-seed is ACCOUNT-ROOT-sensitive. spt-hosted spawn read `brandon@bigscreenvr.com` root where an earlier probe read `Reavo`; doyle's mechanism = detached WMI/schtasks spawn rebuilds env via CreateEnvironmentBlock (registry + token-profile only, drops session-local vars incl. CLAUDE_CONFIG_DIR) → different account root than the operator shell. Pre-trust must target the RESOLVED-SPAWN account root. F-027 seam foundation.
- CreateEnvironmentBlock allow-list MUST include SystemRoot/windir/ComSpec/PATHEXT/PROCESSOR_*/SystemDrive (daemon panics on missing SystemRoot via hickory-resolver) beyond Machine+User regs + token-profile set.
- **rc-offline-after-relaunch anomaly** (seen si1-probe AND b6-probe): a 2nd `endpoint run` for the same id after a tree-kill → rc "offline". Purge + fresh id fixes.

Related: [[v0220-validation-wave]], [[f023-f024-crossnode-delivery]].