docs(benchmark): record Windows-host §11–§14 + plan closeout

§11 Win-guest static (boot 29.75s, +24% vs Linux-host §8; static-IP floor
~22s confirmed host-independent), §12 Lin-guest DHCP (13.51s, +53%), §13
burn-in 4×10 (Win 36/40 — transient WinRM faults under 4× concurrency,
self-recovered; Lin 40/40). §14 adds the full host×guest×IP-mode matrix.
Plan checkboxes filled with results and the template-parity pre-req note.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

docs/RUNBOOK.md

@@ -579,3 +579,188 @@ comparison reflects the *as-deployed* path for each guest, not an
 IP-mode-controlled A/B. The Win guest's static-IP floor (~21.8 s) is dominated
 by Windows boot + Task Scheduler + NIC reconfig; the Linux guest needs no such
 in-guest agent.
+
+---
+
+## 11. Static IP baseline — WinBuild2025 on Windows host (pairs §8)
+
+Recorded 2026-06-07 — `Measure-CIBenchmark.ps1 -StaticIP 192.168.79.200
+-Iterations 10`, **Windows 11 host** + VMware Workstation Pro, template
+`WinBuild2025` / snapshot `BaseClean`, `guestinfo.ip-assignment` injected
+before start; the in-guest `ci-static-ip.ps1` task applies the IP and writes
+back `guestinfo.ci-ip`. IP column = time to `guestinfo.ci-ip` readable
+(`-GuestInfoOnly`). Ready = WinRM/5986 TCP probe. This is the Windows-host
+counterpart to the Linux-host §8.
+
+| Iter    | Clone (s) | Start (s) | IP acquire (s) | Ready (s) | Destroy (s) | Boot total (s) |
+| ------- | --------- | --------- | -------------- | --------- | ----------- | -------------- |
+| 1       | 0.70      | 1.71      | 26.65          | 0.45      | 10.47       | 29.51          |
+| 2       | 0.62      | 1.75      | 24.29          | 2.01      | 24.29       | 28.67          |
+| 3       | 0.63      | 1.76      | 21.83          | 6.01      | 24.32       | 30.23          |
+| 4       | 0.62      | 1.57      | 21.92          | 6.02      | 10.80       | 30.13          |
+| 5       | 0.62      | 1.56      | 24.24          | 2.02      | 24.21       | 28.44          |
+| 6       | 0.62      | 1.55      | 21.94          | 6.01      | 24.28       | 30.12          |
+| 7       | 0.63      | 1.55      | 21.85          | 6.01      | 24.28       | 30.04          |
+| 8       | 0.64      | 1.74      | 21.73          | 6.01      | 11.25       | 30.12          |
+| 9       | 0.62      | 1.57      | 21.90          | 6.01      | 24.28       | 30.10          |
+| 10      | 0.62      | 1.74      | 21.72          | 6.02      | 10.76       | 30.10          |
+| **avg** | **0.63**  | **1.65**  | **22.81**      | **4.66**  | **18.89**   | **29.75**      |
+
+**Host comparison — §8 (Linux host) vs §11 (Windows host), Win guest static:**
+
+| Metric     | §8 Linux host | §11 Windows host | Δ (Win vs Lin)            |
+| ---------- | ------------- | ---------------- | ------------------------- |
+| Clone      | 0.40 s        | 0.63 s           | +58 % (NTFS linked clone) |
+| Start      | 1.88 s        | 1.65 s           | −12 %                     |
+| IP acquire | 21.77 s       | 22.81 s          | +5 % (≈ same)             |
+| Ready      | 0.00 s        | 4.66 s           | +4.7 s                    |
+| Destroy    | 10.40 s       | 18.89 s          | +82 % (bimodal ~11/24 s)  |
+| Boot total | 24.05 s       | 29.75 s          | +24 %                     |
+
+**Key findings**:
+- The **static-IP floor (~21.8 s) is host-independent** — exactly as the plan
+  predicted. It is set by Windows boot + Task Scheduler + NIC reconfig inside
+  the guest, identical across both hosts (Win 22.81 s vs Lin 21.77 s, +5 %).
+- **Ready ≠ 0 on the Windows host** (avg 4.66 s, mostly ~6 s) where it was 0 on
+  Linux. After `ci-static-ip.ps1` rewrites the NIC, WinRM/5986 needs a few
+  seconds to re-bind on the new address before the TCP probe succeeds; on the
+  Linux host WinRM was already listening when `ci-ip` was written.
+- Clone is slower on NTFS (+58 %); destroy is slower and bimodal (~11 s vs
+  ~24 s) — the ~24 s cases hit the 10 s graceful-stop timeout before deleteVM.
+- Net boot-total penalty of the Windows host at constant guest+mode: **+24 %**.
+
+**Pre-req gotcha (template parity)**: the first §11 attempt timed out at 300 s
+every iteration — `guestinfo.ci-ip` never appeared, guest stayed on DHCP.
+Cause: the `F:\CI\Templates\WinBuild2025` `BaseClean` snapshot did **not**
+contain the in-guest `ci-static-ip` agent (added to the Linux-host template
+during B8/§8, never propagated to the Windows-host copy). Host-side guestinfo
+injection works, but with no in-guest consumer the static IP is never applied.
+Fixed by copying the Linux-host template set onto the Windows host (template
+parity), after which all 10 iterations passed. If §11 ever regresses to 300 s
+timeouts, confirm `C:\CI\ci-static-ip.ps1` + the `CI-StaticIp` startup task
+exist inside a clone (see [§13 op-notes](#13-concurrent-capacity-burn-in-on-windows-host-pairs-9)).
+
+---
+
+## 12. Linux guest single-job baseline — Windows host (pairs §10)
+
+Recorded 2026-06-07 — `Measure-CIBenchmark.ps1 -GuestOS Linux -Iterations 10`,
+**Windows 11 host** + VMware Workstation Pro, template `LinuxBuild2404` /
+snapshot `BaseClean-Linux`, **DHCP** (guest publishes its lease via
+`guestinfo.ci-ip`). Ready = SSH/22 probe. Windows-host counterpart to §10.
+
+| Iter    | Clone (s) | Start (s) | IP acquire (s) | Ready (s) | Destroy (s) | Boot total (s) |
+| ------- | --------- | --------- | -------------- | --------- | ----------- | -------------- |
+| 1–10    | 0.62–0.69 | 1.37–1.81 | 11.26–11.42    | 0.00–0.02 | 6.02–6.21   | 13.43–13.83    |
+| **avg** | **0.64**  | **1.50**  | **11.36**      | **0.00**  | **6.13**    | **13.51**      |
+
+**Host comparison — §10 (Linux host) vs §12 (Windows host), Lin guest DHCP:**
+
+| Metric     | §10 Linux host | §12 Windows host | Δ (Win vs Lin)             |
+| ---------- | -------------- | ---------------- | -------------------------- |
+| Clone      | 0.22 s         | 0.64 s           | +191 % (NTFS linked clone) |
+| Start      | 1.14 s         | 1.50 s           | +32 %                      |
+| IP acquire | 7.45 s         | 11.36 s          | +52 %                      |
+| Ready      | 0.00 s         | 0.00 s           | ≈ same                     |
+| Destroy    | 4.94 s         | 6.13 s           | +24 %                      |
+| Boot total | 8.82 s         | 13.51 s          | **+53 %**                  |
+
+**Key findings**:
+- The Linux guest stays near-deterministic on the Windows host too
+  (IP σ ≈ 0.05 s) — no static-IP agent needed; DHCP+guestinfo is stable.
+- The **host penalty is larger for the Linux guest (+53 % boot total) than for
+  the Windows guest (+24 %, §11)** because the Linux guest's boot is so fast
+  (8.82 s) that fixed Windows-host overheads — NTFS clone (+0.4 s) and the
+  vmnet8/NAT DHCP+guestinfo round-trip (+3.9 s on IP-acquire) — are a larger
+  *fraction* of a small total. In absolute terms the host adds ~4.7 s either way.
+- IP-acquire delta (7.45 → 11.36 s) isolates the **host networking effect**
+  (Windows vmnet8 NAT + VMware Tools guestinfo path) on an otherwise identical
+  guest.
+
+---
+
+## 13. Concurrent capacity burn-in on Windows host (pairs §9)
+
+Recorded 2026-06-07 — `Test-CapacityBurnIn.ps1 -Parallelism 4 -Rounds 10`,
+**Windows 11 host** + VMware Workstation Pro, static IP pool
+`192.168.79.201–204` (added to `F:\CI\config.toml` `[ip_pool]` for parity with
+the Linux-host §9). Repo `Simone/burnin-dummy`. Windows guests use the static
+pool; Linux guests skip it (DHCP+guestinfo). Windows-host counterpart to §9.
+
+| Template       | Snapshot          | Build cmd   | Jobs PASS | Rounds OK | Round time (PASS rounds) | Round avg (PASS) |
+| -------------- | ----------------- | ----------- | --------- | --------- | ------------------------ | ---------------- |
+| WinBuild2025   | `BaseClean`       | `build.ps1` | 36 / 40   | 7 / 10    | 88–93 s                  | ~91.6 s          |
+| LinuxBuild2404 | `BaseClean-Linux` | `build.sh`  | 40 / 40   | 10 / 10   | 88–112 s                 | ~96.4 s          |
+
+**Result: PARTIAL** — Linux 40/40 (OVERALL PASS); Windows 36/40 (OVERALL FAIL,
+3 rounds with transient WinRM faults). 76/80 jobs overall.
+
+**Host comparison vs §9 (Linux host):**
+
+| Guest | §9 Linux host    | §13 Windows host             | Δ round avg            |
+| ----- | ---------------- | ---------------------------- | ---------------------- |
+| Win   | 40/40, ~78.6 s   | 36/40, ~91.6 s (PASS rounds) | +16 % + WinRM instability |
+| Lin   | 40/40, ~70.2 s   | 40/40, ~96.4 s               | +37 %                  |
+
+**Key findings**:
+- **Linux-guest concurrency is stable on the Windows host** (40/40, 10/10) just
+  as on the Linux host — SSH transport showed no faults. Round avg +37 % vs §9,
+  consistent with the §12 single-job host penalty.
+- **Windows-guest concurrency is *not* robust on the Windows host.** Rounds 1–5,
+  9–10 were clean (~91 s); rounds 6–8 each lost 1–2 jobs to **transient WinRM
+  faults** under 4× load and ballooned to 269 / 266 / 622 s (30 s connect
+  timeouts + pypsrp retries). Two failure signatures:
+    - `ConnectTimeoutError` to 5986 (WinRM listener briefly unreachable);
+    - `WSManFault 2150858843` "the shell was not found on the server" (the WinRM
+      shell was recycled mid-build).
+  Failures **self-recovered**: each failed job released its IP slot, no clones
+  were orphaned, the IP pool returned to all-free, and rounds 9–10 were clean
+  again. So this is bursty WinRM contention, not a monotonic collapse or a leak.
+- Likely cause: vCPU oversubscription (4 VMs × 4 vCPU = 16 vCPU) plus the higher
+  host-OS overhead of Windows-host VMware tips concurrent WinRM into timeouts;
+  RAM was not the constraint (13.8 GB free of 63.7 GB during the run). The Linux
+  host on the same hardware sustained 40/40 (§9). **Recommendation for sustained
+  4× Windows-guest concurrency on the Windows host**: lower per-VM vCPU, reduce
+  parallelism to 3, or raise WinRM/pypsrp connect timeouts + add a job-level
+  retry. Tracked in `TODO.md`.
+
+**Operational notes (Windows-host benchmark prerequisites)** — discovered while
+running §11–§13; required for any Windows-host orchestration:
+1. **Template parity** — the Windows-host templates must carry the same in-guest
+   agents as the Linux-host templates (`ci-static-ip.ps1` + `CI-StaticIp`
+   startup task for Windows guests; `ci-report-ip` for Linux). Without the
+   static-IP agent, static-IP jobs hang for the full IP timeout (see §11).
+2. **Production venv must be current** — `F:\CI\python\venv` predated the
+   `ip_pool` feature; `[ip_pool]` in config is silently ignored by a stale venv.
+   Re-run `F:\CI\python\venv\Scripts\python.exe -m pip install .` (non-editable)
+   after pulling new code, then verify `import ci_orchestrator.ip_pool` succeeds.
+3. **Set `CI_VENV_PYTHON`** before invoking `Test-CapacityBurnIn.ps1` /
+   `Invoke-CIJob.ps1` interactively. The shim's auto-detect branch dereferences
+   `$IsWindows`, which is undefined under Windows PowerShell 5.1 + `StrictMode`
+   and throws `VariableIsUndefined`; setting `CI_VENV_PYTHON` skips that branch.
+   Production sets it via the runner (`CI_PYTHON_LAUNCHER`). A code fix to guard
+   the `$IsWindows` reference is tracked in `TODO.md`.
+
+---
+
+## 14. Host × guest × IP-mode matrix (complete)
+
+With §11–§13 the Linux-host / Windows-host comparison is symmetric:
+
+| Host \ (Guest, mode)  | Win/DHCP | Win/static | Lin/DHCP | burn-in 4×10                |
+| --------------------- | -------- | ---------- | -------- | --------------------------- |
+| **Linux** (boot/avg)  | §7 102 s | §8 24.1 s  | §10 8.8 s | §9 Win 78.6 s / Lin 70.2 s  |
+| **Windows** (boot/avg)| §6 60.6 s| §11 29.8 s | §12 13.5 s| §13 Win ~91.6 s / Lin 96.4 s |
+
+Read **down a column** = host effect at constant guest+mode; **across a row** =
+guest/IP-mode effect at constant host. Headline: the Windows host adds **+24 %**
+(Win-guest static) to **+53 %** (Lin-guest DHCP) on single-job boot-to-ready,
+the static-IP floor (~22 s) is host-independent, and Windows-guest WinRM is the
+only path that loses jobs under sustained 4× concurrency (Linux/SSH is 40/40 on
+both hosts).
+
+> Note on §6 vs §7: §6 (Win host DHCP) 60.6 s is *faster* than §7 (Lin host
+> DHCP) 102 s, the reverse of every other column. Both are 4-iteration DHCP
+> baselines whose IP-acquire is dominated by high-variance VMware-Tools polling
+> (σ 26–57 s); the 10-iteration static (§8/§11) and Linux (§10/§12) rows are the
+> reliable host-effect signal.

plans/benchmark-windows-host.md

@@ -16,7 +16,7 @@
 ## 1. Cosa esiste già e cosa manca
 
 | RUNBOOK § | Host  | Guest          | IP mode               | Stato        |
-| --------- | ----- | -------------- | ------- | ---------- |
+| --------- | ----- | -------------- | --------------------- | ------------ |
 | §6        | Win   | WinBuild2025   | DHCP                  | ✓ 2026-05-17 |
 | §7        | Linux | WinBuild2025   | DHCP                  | ✓            |
 | §8        | Linux | WinBuild2025   | static                | ✓            |
@@ -26,7 +26,7 @@
 **Mancanti sul Windows host** (target di questo plan):
 
 | Nuovo § | Host | Guest          | IP mode      | Pair-with |
-| ------- | ---- | -------------- | ------- | --------- |
+| ------- | ---- | -------------- | ------------ | --------- |
 | §11     | Win  | WinBuild2025   | static       | §8        |
 | §12     | Win  | LinuxBuild2404 | DHCP         | §10       |
 | §13     | Win  | Win + Linux    | burn-in 4×10 | §9        |
@@ -38,24 +38,23 @@ coerenza (stesso disco/fragmentazione del momento).
 
 ## 2. Prerequisiti sul Windows host
 
-- [ ] Boot in Windows; verificare VMware Workstation + `vmrun.exe` in
+- [x] Boot in Windows; verificare VMware Workstation + `vmrun.exe` in
       `C:\Program Files (x86)\VMware\VMware Workstation\vmrun.exe`.
-- [ ] Template fully powered-off con snapshot integri:
+- [x] Template fully powered-off con snapshot integri:
       ```powershell
       & 'C:\Program Files (x86)\VMware\VMware Workstation\vmrun.exe' -T ws listSnapshots 'F:\CI\Templates\WinBuild2025\WinBuild2025.vmx'
       & 'C:\Program Files (x86)\VMware\VMware Workstation\vmrun.exe' -T ws listSnapshots 'F:\CI\Templates\LinuxBuild2404\LinuxBuild2404.vmx'
       ```
-      Attesi: `BaseClean` e `BaseClean-Linux`.
+      Attesi: `BaseClean` e `BaseClean-Linux`. ✓ (anche `PostInstall` su Win)
-- [ ] Credenziali in **Windows Credential Manager** (non keyring file):
+- [x] Credenziali in **Windows Credential Manager** (non keyring file):
-      target `BuildVMGuest` (guest WinRM) e `GiteaPAT` (clone in-VM).
+      target `BuildVMGuest` (guest WinRM, user `ci_build`) e `GiteaPAT`
-      Verifica: `cmdkey /list | findstr /i "BuildVMGuest GiteaPAT"`.
+      (user `Simone`). Verificato.
-- [ ] Chiave SSH guest Linux presente: `F:\CI\keys\ci_linux` (per §12/§13
+- [x] Chiave SSH guest Linux presente: `F:\CI\keys\ci_linux` (per §12/§13
       lato LinuxBuild2404).
-- [ ] Gitea raggiungibile (`http://10.10.20.11:3100`) e repo
+- [x] Gitea raggiungibile (`http://10.10.20.11:3100`, HTTP 200) e repo
-      `Simone/burnin-dummy` con `build.ps1` + `build.sh` (per §13).
+      `Simone/burnin-dummy` con `build.ps1` + `build.sh` (verificato via API).
-- [ ] `F:\CI\BuildVMs\` privo di cloni orfani e `F:\CI\ip-pool.json`
+- [x] `F:\CI\BuildVMs\` privo di cloni orfani e `F:\CI\ip-pool.json` assente
-      (se esiste) pulito — stesso caveat lease stale del Linux host
+      (pulito). Reset se serve:
-      (RUNBOOK §9 / TODO §2.8). Reset se serve:
       ```powershell
       Set-Content -Path 'F:\CI\ip-pool.json' -Value '{}' -NoNewline
       ```
@@ -71,10 +70,13 @@ coerenza (stesso disco/fragmentazione del momento).
   -StaticIP 192.168.79.200 -Netmask 255.255.255.0 `
   -Iterations 10
 ```
-- [ ] 10 iter completate, tabella per-fase stampata.
+- [x] 10 iter completate, tabella per-fase stampata. **Boot total avg 29.75 s.**
-- [ ] Confronto vs §8 (Linux host static): atteso clone/start simili,
+- [x] Confronto vs §8 (Linux host static): IP-floor confermato host-independent
-      IP-acquire dominato da `ci-static-ip.ps1` (~21.8 s floor, indipendente
+      (22.81 s vs 21.77 s, +5 %). Penalità host: clone +58 % (NTFS), Ready +4.7 s
-      dall'host). Delta host atteso piccolo sulle fasi non-IP.
+      (WinRM re-bind dopo NIC reconfig), destroy +82 % (bimodale), boot +24 %.
+      Vedi RUNBOOK §11. **Nota**: primo tentativo fallito (timeout 300 s/iter) —
+      template Windows privo dell'agent `ci-static-ip`; risolto copiando i
+      template dal Linux host (parità template).
 
 ## 4. §12 — LinuxBuild2404 DHCP (pair §10)
 
@@ -85,10 +87,10 @@ coerenza (stesso disco/fragmentazione del momento).
   -GuestOS Linux `
   -Iterations 10
 ```
-- [ ] 10 iter completate.
+- [x] 10 iter completate. **Boot total avg 13.51 s.**
-- [ ] Confronto vs §10 (Linux host, Lin guest DHCP, boot total 8.82 s):
+- [x] Confronto vs §10 (Linux host, 8.82 s): **+53 %** host-side — clone +191 %
-      isola l'**effetto host** sul guest Linux. Se Win host ≫ 8.82 s,
+      (NTFS), IP-acquire +52 % (NAT vmnet8 Windows + guestinfo). Guest Linux resta
-      l'overhead è host-side (NTFS / NAT vmnet8 Windows), non guest.
+      deterministico (IP σ ≈ 0.05 s). Vedi RUNBOOK §12.
 
 ## 5. §13 — Concurrent burn-in 4×10 (pair §9)
 
@@ -115,9 +117,15 @@ coerenza (stesso disco/fragmentazione del momento).
   -GuestOS Linux `
   -BuildCommand 'bash build.sh'
 ```
-- [ ] Entrambi 40/40 PASS, 10/10 round.
+- [~] **Linux 40/40 PASS (10/10 round). Windows 36/40 (7/10 round).** I 4 fail
-- [ ] Round avg vs §9 (Linux host: Win ~78.6 s, Lin ~70.2 s): delta = effetto
+      Windows = fault WinRM transitori (round 6–8: connect-timeout 5986 +
-      host sotto concorrenza 4×.
+      WSManFault "shell not found") sotto 4× concorrenza; auto-recuperati (slot IP
+      rilasciati, 0 cloni orfani, round 9–10 di nuovo puliti). Linux/SSH 0 fault.
+- [x] Round avg vs §9 (Win ~78.6 s, Lin ~70.2 s): Windows host Win ~91.6 s
+      (round PASS) **+16 %** + instabilità WinRM; Lin 96.4 s **+37 %**.
+      Vedi RUNBOOK §13. **Pre-req scoperti**: (1) parità template, (2) prod venv
+      F:\CI\python\venv stale (pre-`ip_pool`) → reinstallato `pip install .`,
+      (3) `CI_VENV_PYTHON` da settare (bug shim `$IsWindows` su PS 5.1+StrictMode).
 
 ---
 
@@ -125,10 +133,9 @@ coerenza (stesso disco/fragmentazione del momento).
 
 - Tutti i run appendono a `F:\CI\Logs\benchmark.jsonl` (single-job) e
   stampano summary (burn-in).
-- [ ] Copiare/annotare i risultati e aggiungere a `docs/RUNBOOK.md` le
+- [x] Risultati aggiunti a `docs/RUNBOOK.md` §11–§13 (+ §14 matrice completa),
-      sezioni **§11–§13** con le stesse tabelle di §8/§9/§10 e una riga
+      con confronto host Linux-vs-Windows per ciascuna coppia.
-      finale di **confronto host Linux-vs-Windows** per ciascuna coppia.
+- [x] Pulizia post-run: `F:\CI\BuildVMs\` vuoto, artifact `burnin-*`
-- [ ] Pulizia post-run: `F:\CI\BuildVMs\` vuoto, artifact `burnin-*`
       rimossi, `F:\CI\ip-pool.json` reset.
 
 ---