docs: retire the support-vm-1 QA lane

.claude/skills/worldos-dev/SKILL.md

@@ -14,18 +14,16 @@ universe-system that generates worlds. Source-available commercial product, BG-f
 Godot GT2 is archived as reference/extension material under `extensions/renderers/godot/`.
 **AI asset generation** (3D chars, rigging, painterly backdrops, sprites) → the `asset-gen` skill.
 
-> **⚠ HEAVY QA SWEEPS (5-persona / RRI) → USE THE SUPPORT VM, NOT LOCAL.** The 16 GB Mac OOMs mid-sweep
-> (proven 2026-06-02: cratered to 147 M free, personas 2–5 backends never minted, shipped a junk PARTIAL RRI 2.7).
-> Documented lane: **`WorldOS-GUI-RUNBOOK.md` § "Support VM lane (heavy sweeps)"** + memory note
-> `project_worldos_vm_sweep_lane.md`. The Mac runs the **macOS-only** native Part-A (#356 `.app` launcher +
-> built-app screenshots); the **32 GB evaos-support VM** (`root@178.104.123.213`, repo `/root/worldos-qa/WorldOS`,
-> harness `sweep_v2.sh`) runs the heavy **part-B** 5-persona sweep against the **real OpenWorlds browser GUI**
-> (server.py + playwright/chromium — NOT a degraded proxy). RRI rolls up Mac `--handoff-json` + same-SHA VM dirs.
-> **On resume after a compaction, re-read that runbook section FIRST** — do not rebuild the QA plan from memory
-> or fight local RAM. For per-beat DM latency, see the `worldos-latency-forensics` skill.
-
-## VM GATE SWEEP — exact procedure (heavy part-B; the thing you keep re-deriving)
-**The full gate is ONE command on the 32 GB VM. Update THIS section when the harness changes** — it is
+> **Heavy QA runs on the local Mac.** The support VM lane (support-vm-1) was retired on 2026-10-30. See evaos-support-control#1109.
+> The release part-B 5-persona sweep has no local port yet: its harness was VM-only (#1822).
+> (Historical: the old 16 GB Mac OOMed mid-sweep on 2026-06-02, which is why the VM lane existed. The Mac now
+> has 14 cores / 64 GB.)
+> For per-beat DM latency, see the `worldos-latency-forensics` skill.
+
+## VM GATE SWEEP — historical procedure (retired 2026-10-30 — do not follow)
+
+**The support VM lane (support-vm-1) was retired on 2026-10-30; local port of this sweep: #1822. See evaos-support-control#1109.** Kept as a record only.
+The full gate was ONE command on the 32 GB VM — it was
 the source of truth so a cold post-compaction agent never re-learns it (it cost a real diagnosis 2026-06-03).
 
 **The VM** (`evaos-support`): `root@178.104.123.213`, key `~/.openclaw/secrets/cloud-deploy-key` (key-only

.claude/skills/worldos-latency-forensics/SKILL.md

@@ -91,9 +91,9 @@ between tool calls" read — same conclusion, correct metric.)
 2. **Cache-stability** (one 2-beat run) → confirm the prefix cache holds across beats.
 3. **Effort/flag-wiring probe** → confirm the runner ACTUALLY consumes the flag (see worldos-dev's
    "QA must exercise the flag" — a harness that ignores `WORLDOS_LEAN_BEATS`/effort fakes the A/B).
-4. **Short duo A/B on the VM** (engine-duo, gateway-free) → quality + latency deltas, same SHA + seed,
+4. **Short duo A/B on the local Mac** (engine-duo, gateway-free) → quality + latency deltas, same SHA + seed,
    one credit window (throttling confounds wall-clock).
-5. **ONLY THEN** the 5-persona `.app` gate (heavy — use the Support VM lane; see worldos-dev).
+5. **ONLY THEN** the 5-persona `.app` gate (heavy; the support VM lane that ran its part-B sweep was retired 2026-10-30, local port in #1822).
 Record latency columns (s/beat, cold-open-s, turns/beat) to `qa/scores_db.py` on every run.
 
 Cross-refs: `worldos-dev` (dev/QA loops + VM lane), `first-principles-architectural-decision`

AGENTS.md

@@ -11,7 +11,7 @@ Bootstrap order: docs/OPERATIONS.md → docs/roadmap/NOW.md → docs/ACTIVE-GOAL
 - Use `/Users/m1/Codex` for Codex artifacts, scratch files, screenshots, reports, and downloaded CI/VM artifacts.
 - Use same-disk local worktrees for GUI/native-app edits that must launch against private art. Lexar worktrees are fine for docs, backend-only, and non-GUI slices that do not launch the app against private art.
 - Before running install, build, or test commands, verify `pwd`. If a GUI/native app run is not in `/Users/m1/WorldOS` or a same-disk worktree with `WORLDOS_ART_REPO_ROOT=/Users/m1/WorldOS`, explain why.
-- Prefer the local Mac (14 cores / 64 GB) for heavyweight validation, full suites, matrix tests, long integration tests, and persona sweeps; use GitHub Actions when a remote runner is better suited. `support-vm-1` is a customer box, not the default heavy-sweep host.
+- Prefer the local Mac (14 cores / 64 GB) for heavyweight validation, full suites, matrix tests, long integration tests, and persona sweeps; use GitHub Actions when a remote runner is better suited. The support VM lane (support-vm-1) was retired on 2026-10-30; use the local Mac lane. See evaos-support-control#1109.
 - Run local tests only for fast feedback, local-only reproduction, validating unpushed edits, or Mac-only `.app` proof. Use the narrowest focused command first.
 - Do not launch multiple heavyweight local suites or persona sweeps in parallel on this Mac.
 - If local test work causes memory pressure, stop it, report the command/path, and switch to a narrower check or GitHub CI.
@@ -25,14 +25,12 @@ Bootstrap order: docs/OPERATIONS.md → docs/roadmap/NOW.md → docs/ACTIVE-GOAL
 - The current `qa/RRI.json` from `f5500ac` is partial/harness-contaminated evidence, not a release verdict.
 - Release evidence requires the RRI contract in `qa/release_readiness.py`: expected/completed/missing personas, disk-backed scores, behavior/UI/image/palette evidence, same build SHA, and non-partial status.
 
-## Support VM
+## Support VM (retired 2026-10-30)
 
-- Target VM: owner-provided 32GB support VM, `support-vm-1`.
-- Connection/auth details are operator-only and should stay outside tracked repo docs.
-- Use `support-vm-1` only for explicitly customer-box or remote-runner work after its credentials/config are intentionally installed and verified there; local heavy QA remains the default.
-- VM preflight must record VM identity, repo checkout path, branch/SHA, Codex CLI version, auth/profile status, `uv`, Node/npm/Playwright availability, private-art status or explicit backend-only/no-art classification, env vars, budget/concurrency cap, teardown commands, and artifact return path under `/Users/m1/Codex`.
-- The VM cannot prove Mac-only surfaces. `WorldOS.app` build/launch, native #356, and built-app UI play evidence stay on this Mac or macOS CI.
-- VM artifacts can feed RRI only when `run.json`, `score.json`, `session_surface.final.json`, network/image evidence, palette-live evidence, and build SHA are explicit. Otherwise the result remains partial/harness-contaminated.
+- The support VM lane (support-vm-1) was retired on 2026-10-30; use the local Mac lane. See evaos-support-control#1109.
+- Historical rules, kept for audit context: VM preflight had to record VM identity, repo checkout path, branch/SHA, Codex CLI version, auth/profile status, `uv`, Node/npm/Playwright availability, private-art status or explicit backend-only/no-art classification, env vars, budget/concurrency cap, teardown commands, and artifact return path under `/Users/m1/Codex`.
+- The VM could not prove Mac-only surfaces. `WorldOS.app` build/launch, native #356, and built-app UI play evidence stay on this Mac or macOS CI.
+- VM artifacts could feed RRI only when `run.json`, `score.json`, `session_surface.final.json`, network/image evidence, palette-live evidence, and build SHA are explicit. Otherwise the result remains partial/harness-contaminated.
 
 ## GEX44 GPU host (HISTORICAL — retired 2026-08-06)
 
@@ -42,7 +40,7 @@ build, capture, or save against `46.4.26.123`, `/home/unity`, or the retired `ge
 Unity 6000.5.6f1 runs at
 `/Users/m1/worldos-unity` (mirror `/Users/m1/Codex/worldos-unity-mirror`) via
 `extensions/renderers/unity/tools/mcp_stdio_exec.py`; build with the WorldOS macOS menu command and run
-QA through `qa/qa_sandbox.py`. Local Mac heavy QA is primary; `support-vm-1` is a customer box.
+QA through `qa/qa_sandbox.py`. Local Mac heavy QA is primary; `support-vm-1` was retired 2026-10-30.
 
 ## Shared Owned-Repo Policy
 

VISION.md

@@ -156,7 +156,7 @@ Tier-0/1 result as a release verdict):
   rotated persona + a ≥6-beat duo (below 6 disarms the FATAL behavioral floors). Iteration signal only.
 - **Tier 2 — the 5-persona sweep + Mac native part-A → RRI** (~$40, ~90min): milestone verdict
   only, after 0+1 pass; ≤3 sweeps per milestone. Heavy QA runs locally on this 14-core / 64 GB Mac;
-  `support-vm-1` is a customer box and is not the default sweep host. Log every scored run to the
+  `support-vm-1` was retired on 2026-10-30. Log every scored run to the
   ledger (`qa/scores_db.py`).
 
 **Release ladder:** Engine Excellent → **Player-Ready Beta** (a real built `.app`; a no-prior-

WorldOS-GUI-RUNBOOK.md

@@ -248,7 +248,7 @@ evidence gaps can count as release evidence.
   4/5 derived — a self-report-coverage artifact, not a quality verdict). Read `satisfaction_source`; a derived
   G3-miss is inconclusive.
 - **PARTIAL ≠ product RED.** A host-memory crater (only persona-1 scored) or a hung/incomplete rollup (e.g. the
-  duo step stalls) is harness-contaminated — re-run via the VM lane; never cite a PARTIAL RRI as a quality score.
+  duo step stalls) is harness-contaminated — re-run it on a clean lane (the VM lane was retired 2026-10-30; local port in #1822); never cite a PARTIAL RRI as a quality score.
 
 ## macOS privacy prompt triage
 
@@ -277,15 +277,11 @@ script/build_and_run.sh --verify
 This proves the local/worktree-built bundle launches without killing an existing app. It is only a smoke:
 release truth still requires `qa/ui_playtest_app.sh` Part A+B and the full RRI sweep.
 
-## Support VM lane (customer-box / explicitly remote; not the default)
+## Historical support VM lane (retired 2026-10-30 — do not follow)
 
-> **Local is primary:** this 14-core / 64 GB Mac owns Unity, visual rendering, and heavy QA. The
-> 32 GB `support-vm-1` is a customer box; use it only for an explicitly remote/customer-box run.
-> This section records that lane's mechanics and is not a reason to move default heavy sweeps off Mac.
+> **The support VM lane (support-vm-1) was retired on 2026-10-30; use the local Mac lane. See evaos-support-control#1109.** The host was wiped and repurposed; it no longer has WorldOS tooling, credentials, the
+> Woodpecker agent, or tailnet access. The notes below are kept only as a dated record.
 
-- Target: owner-provided **32GB support VM** (`support-vm-1`); connection: `root@178.104.123.213`,
-  key `~/.openclaw/secrets/cloud-deploy-key`, repo `/root/worldos-qa/WorldOS`. Further connection/auth
-  detail lives in local operator-only runbooks/evidence, not tracked repo docs.
 - Do not assume it is ready for an explicitly remote/customer-box run until credentials/config are intentionally installed and verified.
   When selected, the support-VM persona lane is Codex DM plus Codex UI player; Claude is only required when
   the preflight is run with `--provider claude` or `--player-agent claude`. The Codex lane requires
@@ -374,7 +370,7 @@ run QA via `qa/qa_sandbox.py` (8866/8972; owner 8776/8981). Keep source authorit
 
 The **Hetzner GEX44 GPU dedicated server** (`evaos-gpu-gex44-1`, provisioned 2026-06-25; RTX 4000 SFF Ada /
 64 GB / Ubuntu 24.04) was the former heavy-QA and Unity render host. This dated record is retained for
-evidence provenance only; use the Local Unity lane and `support-vm-1` above for current work.
+evidence provenance only; use the Local Unity lane above for current work (`support-vm-1` was retired 2026-10-30).
 **Connection/auth details live in `~/.openclaw/secrets/gex44.env` (operator-only)** — never put the endpoint,
 SSH key, or the VNC password in this tracked doc (same convention as the Support VM lane).
 
@@ -546,8 +542,8 @@ took over the developer's only display and the Mac had to be rebooted. Now:
   (`worldos-*`/`WORLDOS_*` MCP ids, `dev.worldos.app`); you MAY read `WORLDOS_ART_REPO_ROOT`.
 - `_private/` (the 2.9 GB art) is **never committed**. Building/serving from the local checkout is how the
   art is present; worktrees can read it via `WORLDOS_ART_REPO_ROOT=/Users/m1/WorldOS` when needed.
-- 14-core / 64 GB Mac: local heavy QA is the default. Use GitHub CI or the customer `support-vm-1`
-  only when the run is explicitly remote; never launch multiple heavyweight persona sweeps locally.
+- 14-core / 64 GB Mac: local heavy QA is the default. Use GitHub CI only when the run is explicitly
+  remote (`support-vm-1` was retired 2026-10-30); never launch multiple heavyweight persona sweeps locally.
 - **Verify, don't trust:** ≥2 clean reads for any claim; the RRI scorer reads disk, not the live
   channel; confirm builder PRs actually pushed.
 - The product is the **launchable, played .app**. A green score on any other surface is a

WorldOS-OPERATING-GOAL.md

@@ -24,7 +24,8 @@
      WORKTREES:    For tracked edits, prefer same-disk local worktrees when GUI/app tests need assets.
                    Use /Users/m1/Codex for evidence/snapshots and Lexar worktrees only for
                    non-GUI/doc/backend slices that do not launch the app against private art.
-     SUPPORT VM:   32GB owner-provided support VM (`support-vm-1`). Connection/auth details live
+     SUPPORT VM:   RETIRED 2026-10-30 (support-vm-1); use the local Mac lane. See
+                   evaos-support-control#1109. HISTORICAL RECORD ONLY, do not follow: 32GB support VM. Connection/auth details lived
                    in local operator-only evidence/runbooks, not tracked repo docs. Use it for
                    heavy backend/persona sweeps only after Codex/config/credentials are intentionally
                    installed. The default VM persona lane is Codex DM plus Codex UI player and requires
@@ -69,7 +70,8 @@
      LAST VALID RELEASE GATE:
                    none after the RRI contract hardening. A release verdict requires expected
                    persona count, disk-backed palette/image/behavioral evidence, and built .app play.
-     NEXT ACTION:  RE-MEASURE — the binding step. Run the HARDENED sequential VM sweep (#844: quota-safe,
+     NEXT ACTION:  RE-MEASURE — the binding step. First land the local part-B sweep (#1822; the VM
+                   lane was retired 2026-10-30), then run the HARDENED sequential sweep (#844: quota-safe,
                    429-aborts, ABORTED-not-a-number) at CURRENT main + a same-SHA Mac part-A handoff →
                    full rollup (--handoff-json + --support-preflight-json + --abort-marker) → ruler-stamped
                    row (run_id v1.0.4-rc4-<sha>; lens auto-stamps; ONE branch — scores.db is a committed
@@ -296,10 +298,10 @@ verifier; can revert the goal to "fix" anytime.
   `/usr/bin/find` launched by the test/diagnostic environment. Treat that screenshot prompt as harness
   contamination unless a clean run shows `WorldOSApp`/WebKit itself accessing a protected library path.
 - The next gate evidence step is issue #466: a clean non-partial five-persona RRI from one explicit SHA.
-  The easiest current path is a support-VM sweep pinned to `da05101`, gated first by
+  It needs a local part-B sweep first (#1822); the support VM `support-vm-1` was retired 2026-10-30.
+  (Historical, do not follow: the old path was a support-VM sweep pinned to `da05101`, gated first by
   `qa/support_vm_preflight.py --expected-sha da05101`, and paired with the `da05101` handoff JSON. If the
-  sweep runs on a newer `origin/main` tip, rerun the Mac handoff on that same SHA first.
-  Heavy backend/persona sweeps belong on the owner-provided 32GB support VM (`support-vm-1`) once auth/config
+  sweep ran on a newer `origin/main` tip, the Mac handoff was rerun on that same SHA first.) Heavy backend/persona sweeps belonged on the 32GB support VM once auth/config
   are intentionally installed there; connection details are kept outside tracked docs. The 2026-06-07
   repo-owned support preflight for `da05101` wrote
   `/Volumes/LEXAR/Codex/worldos-support-vm-rri/da05101-preflight/` (historical path — /Volumes/LEXAR/Codex/worldos-support-vm-rri/da05101-preflight/; that machine is gone, see NOW.md) and returned `verdict=blocked`:

WorldOS-RUNBOOK.md

@@ -12,11 +12,8 @@
 > **Local/VM routing, 2026-06-01:** `/Users/m1/WorldOS` is the synced local app/private-art
 > checkout and should be used for GUI/native-app testing. Use `/Users/m1/Codex` for evidence,
 > snapshots, and logs; do not make Lexar the default GUI runtime tree because external-drive
-> permissions can break local AI/browser tests. Heavy backend/persona sweeps belong on GitHub CI or
-> the owner-provided 32GB support VM (`support-vm-1`) after remote access and Codex config are
-> intentionally installed and verified; connection details are kept outside tracked docs. A read-only
-> scout reached the operator endpoint and found `evaos-support` suitable but stale (`4524b3e`, behind
-> the `9545383` proof baseline) with Codex auth/config unproven. Mac-only built-app proof remains local/macOS CI.
+> permissions can break local AI/browser tests. Heavy backend/persona sweeps run on this Mac or GitHub CI.
+> The support VM lane (support-vm-1) was retired on 2026-10-30; use the local Mac lane. See evaos-support-control#1109. Mac-only built-app proof remains local/macOS CI.
 
 > **This is the compaction-resilience doc.** If you are an agent resuming this project
 > after a context reset, read this top-to-bottom before doing anything. It captures the

docs/KIMI-ONBOARDING.md

@@ -64,7 +64,7 @@ Source: `~/.claude/tasks/237280f0-e8fe-4529-bb6e-72957d537c61/76.json` (status:
   through `extensions/renderers/unity/tools/mcp_stdio_exec.py`; build with
   `execute_menu_item "Tools/WorldOS/Build/macOS Player (Universal)"`,
   capture with `manage_camera`, and run QA via `qa/qa_sandbox.py` (8866/8972; owner 8776/8981).
-- Commit renderer source here; `support-vm-1` is only the heavy-backend-sweep fallback.
+- Commit renderer source here; `support-vm-1` was retired on 2026-10-30; heavy sweeps run on the local Mac.
 
 ## 6. Skills & agent profiles — porting decision (read-on-demand, NO symlinks/copies)
 

docs/MODEL-TIERING-STRATEGY.md

@@ -126,8 +126,8 @@ artifact (little combat to score), not an engine defect (Engine-Excellent is met
 ## Validation ladder (cheap → expensive; before any model/effort spend)
 digest-correctness (1 engine call, no LLM) → cache-stability (1 two-beat run) → effort/flag-wiring probe
 (confirm the runner consumes the flag — see worldos-dev "QA must exercise the flag") → short duo A/B on the
-32 GB VM (same SHA + seed, one credit window) → ONLY THEN the 5-persona `.app` gate (Mac native Part-A + the
-VM part-B sweep; see the Support VM lane). Hold the scorer constant during any A/B.
+local Mac (same SHA + seed, one credit window) → ONLY THEN the 5-persona `.app` gate (Mac native Part-A + the
+part-B persona sweep; the Support VM lane that ran part-B was retired 2026-10-30, local port in #1822). Hold the scorer constant during any A/B.
 
 ## Still open
 - **The #640 lean re-ground root** — the key Opus-latency enabler (re-enable lean safely; repro-first).