memory: snapshot — run 7 training on gx10; run 6 TRANSFERRED after CSAM adjudication; erp-tune-v6-nvfp4a16 live as trial; ESH/YTVC/webhook repairs; ana-ml2 routes persisted; tank/zroot actions deferred to next session. Index 830→271 lines: 27 decisions + 8 abandoned archived, superseded in-flight blocks archived verbatim

This commit is contained in:
vh
2026-09-09 00:26:44 -07:00
parent 3e18a044bd
commit 5ad948bf31
23 changed files with 2951 additions and 2800 deletions
@@ -1,192 +0,0 @@
# DFlash2 speculative decoding — measured on our own stack (2026-08-22)
Operator-driven session. **Read the epistemic labels.** During the chase we generalised from
observations that later proved wrong; this file separates what was *measured* from what remains
*hypothesis*, and records the wrong turns so nobody re-derives them.
## What DFlash2 is
A **2B draft model** (3.85 GB bf16) for speculative decoding against Qwen3.8-27B —
`incoai/Qwen3.8-27B-DFlash2`, Apache-2.0, blog `inco.ai/blog/dflash2`, upstream `z-lab/dflash`.
Block diffusion: drafts a whole 8-token block in one pass, with a candidate selector tracing a
path through per-slot top-K. Lossless (greedy matches the target).
vLLM support merged **2026-08-21 05:27 UTC** as PR **#52816** (`b389ac29`). Method string is
**`"dflash"`**, not `dflash2`.
## ✅ MEASURED — throughput and acceptance
Single instrument (`specbench.py`, 8 fixed prompts, temp 0, max_tokens 256), delta against
vLLM's own `spec_decode` counters. The MTP k=3 numbers reproduce our recorded 58.4% / 55.3%
figures exactly, which is what validates the instrument.
| seat | config | accepted tok/forward | throughput |
|---|---|---|---|
| gen (orcarouter) | MTP k=3 *(production)* | 2.753 | 114.9 tok/s |
| gen | MTP k=7 *(control)* | 3.041 | **74.0 tok/s** |
| gen | **DFlash2 k=7** | **3.254** | **131.9 tok/s** |
| sec (M.O.G.-SEC) | MTP k=3 *(production)* | 2.676 | 110.5 tok/s |
| sec | **DFlash2 k=7** | **3.252** | **130.0 tok/s** |
**⭐ The k=7 MTP control was essential and inverted the obvious read.** Going deeper on MTP
*improves acceptance* (2.753 → 3.041) while **destroying throughput** (114.9 → 74.0). Our MTP
head is a single module (`mtp_num_hidden_layers=1`, only `mtp.layers.0`, 15 tensors) run
autoregressively, so k draft tokens cost k sequential forward passes. **"Just raise
num_speculative_tokens" is a trap** — without the control I would have recommended it.
DFlash2's win is therefore **not better per-token acceptance** — our MTP is actually *better* at
position 0 (79.6% vs 75.4%). It is that block drafting makes depth nearly free.
**⭐ The drafter is model-agnostic across finetunes — 3.254 (gen) vs 3.252 (sec), a 0.06%
difference**, with superimposable per-position curves. One drafter file on `/tank` serves both.
## ✅ MEASURED — how DFlash2 runs (answers "can one drafter serve both seats?")
**EAGLE3-style coupled, not standalone.** In vLLM: `load_model(self, target_model)` binds it to a
specific target object; `pass_hidden_states_to_model=True`; `gpu_model_runner` reads
`dflash_config.target_layer_ids` → `[i+1 …]` to register auxiliary hidden-state capture on the
target at layers **5, 19, 33, 47, 61**. It even reads the target's RoPE style at load.
Consequences:
- **Weights file is shareable** (one download, both seats mount it) — gen and sec are
architecturally identical on every dimension the drafter needs: 64 layers (deepest tap 61),
hidden 5120, intermediate 17408, vocab 248,320 > mask token 248,070.
- **VRAM is NOT shareable — 3.85 GB per seat.** The drafter lives inside the target's engine
process, consuming hidden states mid-forward. Two seats are two processes; there is no
cross-process sharing mechanism and there could not be.
## ✅ MEASURED — it works on our stack, which the card does not claim
The card tests stock BF16 on an H200 with FlashAttention 3. Verified here instead:
**abliterated + NVFP4 `compressed-tensors` target ✓, Blackwell sm_120 ✓, DFlash2 CUDA graphs
captured ✓.** None of that was documented anywhere.
## 🔶 HYPOTHESIS — why our acceptance trails the published numbers
Both our targets land at ~3.25 accepted length against the card's 4.10–5.46 on stock BF16.
**Finetune drift is ruled out** — two *different* finetunes gave identical results to three
decimals. The shared variable is **NVFP4 quantization of the target**, which is mechanically
plausible (the drafter reads quantized hidden states at its five taps). Second candidate:
prompt distribution (ours general-purpose, theirs GSM8K/MATH/HumanEval/MBPP/MT-Bench).
**Neither is confirmed.** Settling it needs a BF16 target seat (~56 GB) — a real GPU window.
## ❌ RETRACTED — the "MTP head mismatch causes the degeneration" hypothesis
**Operator ruling, 2026-08-22: this hypothesis is WRONG. The degeneration lives in the un-fixed
vLLM, not in the weights.** Recorded here rather than deleted, because it was reasoned to
confidently enough that a future session could re-derive it.
**Two independent failures produced it, and the second is the instructive one:**
1. **I treated a false dichotomy as a deduction.** Having verified gen and sec run an identical
engine (same image ID `sha256:bd3236cff208…`, same live version
`0.27.2rc1.dev150+g311b3513a` read from inside both processes, same flags bar
`gpu-memory-utilization` 0.43 vs 0.44), I concluded "config is eliminated, therefore it is the
weights." That does not follow. **An engine bug present in BOTH seats is not exonerated by the
two seats being identical** — it just means the engine cannot explain a *difference*. It can
still explain the *failure*.
2. **The difference I was explaining may not exist.** The premise was a single operator
observation of sec degenerating at ~2k, made during a session with many concurrent changes.
**n=1 under heavy concurrent modification is not evidence** — see the meta-lesson below.
**What survives as fact** (measured, still true, just not causal): sec's MTP head *is*
byte-identical to `qwen38-27b-uncensored-bf16` across all 15 tensors — a stock head on a
security-finetuned body, because the `Qwen3_5ForConditionalGeneration` wrapper never loads the
head, so the finetuning could not reach it. gen's orcarouter head *was* abliterated in-band by
its author. Acceptance differs slightly (gen 58.4%, sec 55.9%). **All true. None of it shown to
cause multi-turn degeneration.**
**Current standing explanation: the degeneration is an engine bug in the un-fixed vLLM.** Both
production seats run `311b3513`, which is **172 commits behind GDN spec-decode fix #53077**
(merged 2026-08-20). `#51113` is present in that build and is therefore **necessary but
insufficient** on its own.
## ⭐⭐ META-LESSON — n=1 during a busy session is not evidence
The operator's own framing, and it generalises past this incident: **an observation made while
many things are being changed at once cannot carry a causal claim, no matter how confidently it
is reported.** Tonight that single observation became the load-bearing premise for a weights-side
hypothesis, a root-cause narrative, and very nearly a recommendation.
This is the same failure the gen-seat compose file already warns about in different words — *"a
passing probe is NOT sufficient evidence"* — inverted. That note guards against trusting a
**negative** result from a synthetic test. This one guards against trusting a **positive**
sighting from an uncontrolled session. Both reduce to: **hold the system still, or do not draw
causal conclusions from it.**
Applies equally to the "coherent to 10k" observation below — same n, same conditions, opposite
direction. Neither observation is worth more than the other.
## ⚠️ CONFOUNDED — and the "before" state is itself unreliable
sec now runs DFlash2 on a newer build and the operator reports **coherent to 10k tokens with
adversarial nonsense prompts**. ⚠ Treat this the same way as the 2k sighting it is being compared
against: **n=1, uncontrolled session, not evidence.** The comparison is weak on *both* ends.
**Two variables changed at once:**
1. **Engine**: `311b3513` → `e9d1398d`, **+259 commits, `behind_by=0`** (a strict superset),
including GDN spec-decode fix **#53077** (merged 2026-08-20) that production is **172 commits
behind**.
2. **Drafter**: frozen MTP head → DFlash2 reading live hidden states.
**Isolating it = run MTP k=3 on the same new build.** Not yet done.
**#51113 is present in BOTH builds** (verified by ancestry, `behind_by=0` each) — so the
"proper upstream fix" our compose comment credits is **necessary but insufficient**; sec ran it
and still degenerated. Related open upstream: **#53180** (quantized Qwen3.8-27B hybrid GDN + MTP
producing *silent* degenerate output, no fix), **#41884** (DFlash + prefix caching on hybrid,
IndexError, workaround is disabling one).
## ❌ WRONG TURNS — do not repeat
- **Version strings are not lineage.** The DFlash2 build reports `0.26.1rc1.dev1048` and our
production nightly `0.27.2rc1.dev150`, which *looks* like a regression. It is a setuptools_scm
tag-reachability artifact. **Use the GitHub compare API and check `behind_by`.**
- **Docker Hub push timestamps lie about source freshness.** `nightly-ba07e4a4` was *pushed*
06:12 UTC, comfortably after the 05:27 merge — but *cut* from a 03:46 commit that predates it.
**Grep the image for the symbols you need.** Believing the timestamp would have cost an RP-seat
outage to serve a model the engine could not instantiate.
- **`--max-num-batched-tokens` was not the image truncation.** Raising it 16,384 → 32,768 on that
theory changed nothing and cost ~3 GiB of peak activation, which came straight out of the KV
pool. The cap was the tokenizer (§3.14 of the playbook).
- **"1M needs YaRN, absent from config" is FALSE for the sec quant.** It is fully present:
`rope_type: yarn`, `factor: 4.0`, `original_max_position_embeddings: 262144`,
`max_position_embeddings: 1000000`. Context is a KV-memory choice, not a model limit.
## Live state — PROMOTED to the compose stack 2026-08-22
**Operator-approved after real-use testing** ("performing very well"). The experimental
standalone container is gone; `stacks/mog-sec/` is canonical and `restart: unless-stopped` means
it survives reboots. Cutover verified: **KV pool 526,617 / 1.10x — identical to the container it
replaced**, restarts 0, both gateway aliases serving, DFlash2 confirmed drafting at k=7
(231 draft tokens over 33 drafts), vision working.
⚠ **One variable was deliberately REMOVED, not carried over.** The old stack hardcoded
`PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True`; the validated DFlash2 container never set it,
and playbook §3.10 records expandable_segments corrupting retained tensors elsewhere. The compose
now defaults it EMPTY (`MOG_ALLOC_CONF`). Promoting it as-was would have shipped a variable the
tested configuration did not have.
**Compose is now parameterised for the shapes that differ:** `MOG_SPEC_CONFIG` carries the whole
speculative JSON (dflash needs `"model": "/drafter"`, MTP must not have one — a method+tokens
template cannot express both), plus `MOG_MM_PROCESSOR_KWARGS`, `MOG_DRAFT_MODEL`,
`MOG_MAX_NUM_BATCHED_TOKENS`, `MOG_ALLOC_CONF`.
**ROLLBACK:** `.env.bak-pre-dflash2-20260822` and `compose.yaml.bak-pre-dflash2-20260822` on the
host; or one line — `MOG_SPEC_CONFIG={"method": "qwen3_5_mtp", "num_speculative_tokens": 3}` plus
the old `MOG_IMAGE`.
| | production sec | current |
|---|---|---|
| image | `nightly-311b3513` | `nightly-e9d1398d` |
| speculation | MTP k=3 | **DFlash2 k=7**, drafter `/tank/aimodels/qwen38-27b-dflash2-drafter` |
| max-model-len | 262,144 | **480,000** |
| KV pool | 418,218 (1.60×) | **526,617 (1.10×)** |
| images | 4096² → 16,384 tok | **2048² → ~5,125 tok** (`--mm-processor-kwargs` size cap) |
⚠ **`--gpu-memory-utilization 0.55` is the stable ceiling** while GPU1's other tenants are up.
0.58 sized KV at 594,172 then **OOM'd during CUDA graph capture** — the process reached 57.49 GiB
against ~57.6 free. Real 1M context needs ~49 GiB of KV and therefore evicting most of GPU1.
Canonical config: `stacks/mog-sec/{compose.yaml,.env.example}` in this repo.
@@ -1,34 +0,0 @@
# [2026-08-23] Every secret-bearing `.env` on ana-docker tightened to 0600
Found while taking uptime ownership of hrafn: its `.env` was mode 0644 with a live
bearer token. Not a hrafn lapse — **0644 was the de facto pattern on the host**.
Eight stacks carried secret-shaped vars in world-readable `.env` files on a box with
four interactive accounts, verified as real exposure by reading one as `nobody`.
Swept: **vaultwarden, traefik**, beszel, gitea-runner, miniflux, news-digest,
searxng, vor. (hrafn and nevermore were fixed separately the same day.) Six other
stacks already used 0600, so this converged on the existing house pattern rather
than inventing one. Post-sweep the host has **zero** secret-bearing `.env` readable
by `nobody`.
Playbook: `playbooks/tighten-env-perms.yaml`, one run per stack, re-runnable.
## The check that matters
Every run asserts `docker compose config` still renders **as the deploy user**
(`lkraven`), not as root. Checking the mode proves the bits changed; only rendering
as the deploy user proves the next deploy can still resolve its variables.
## Two gotchas recorded in the playbook
- **vaultwarden looked like it bind-mounted its `.env`** — which would mean the
*container's* UID reads it and 0600 could break the password vault. It does not:
that `- .env` is under `env_file:`, not `volumes:`. My grep matched the YAML list
item without checking its parent key. The playbook now **refuses** any stack that
genuinely bind-mounts its `.env`, since that case is read by the container UID.
- **elway prompted for a sudo password.** The `ana-docker` ssh alias resolves to
`lkraven`, who needs one; **`infra-ops@10.250.50.70` has NOPASSWD**. `corviduo-dev`
was repointed to infra-ops at some point and `ana-docker` was not. Run elway against
the infra-ops target on this host.
Commit `a896c0a`.
@@ -1,663 +0,0 @@
# [2026-08-23] Anaheim's IPsec tunnel delivers ~25% of a verified 2 Gbps circuit
> **⛔ SUPERSEDED 2026-08-23 (same day, later session) — read the CORRECTION at
> the bottom before acting on anything here.** The headline is wrong (the
> relevant ceiling is NH3's **1 Gbps** uplink, not Anaheim's 2 Gbps), the
> aggregate number is wrong (**692 Mbit/s** at 8 streams, not ~550 — the
> original stopped measuring at 4), and the proposed remedy is **impossible**:
> UniFi's manual site-to-site IPsec does not implement AES-GCM at all. The
> per-stream observation and the parallelise-your-transfers mitigation are the
> parts that survive.
The operator noticed site-to-site transfers were slow for a datacenter fiber
handoff and asked whether WireGuard was the limit. It is not WireGuard, and the
circuit is fine.
## Measured
```
ana-docker -> internet, 8 parallel 2,153 Mbit/s <- the 2 Gbps handoff, delivering
ANA <-> NH3 through the tunnel, 4 par. 460 Mbit/s
FortiGate's own recorded peak 554 Mbit/s
ANA <-> NH3, single stream 227 Mbit/s
ANA <-> ESH, single stream 249-265 Mbit/s
ESH <-> NH3 (never touches ana-gw) 545-557 Mbit/s on a SINGLE stream
```
Method: stdlib TCP probe (no ssh, no crypto, no compression) between site
endpoints; raw circuit measured with 8 parallel HTTPS fetches from Hetzner
Ashburn. Host NICs are virtio with no reported cap, so no host-side ceiling.
## What it is not
- **Not WireGuard.** Both Anaheim tunnels are IPsec on ana-gw
(`pfi-ana-nh3` -> 70.230.226.88, `ana-eshudm-dyn` -> the ESH UDM). WireGuard
on ana-wg is remote-access only and is not in this path. Traceroute confirms:
both slow paths have hop 1 = `10.250.50.1` (the FortiGate); the fast
ESH<->NH3 path rides a `192.168.x` Site Magic overlay and never touches it.
- **Not CPU or crypto exhaustion.** FortiGate CPU was **100% idle across all
8 cores** during the tests, and both live tunnels report `npu_flag=03` with
`dec_npuid=1 enc_npuid=1` — encrypt *and* decrypt are hardware-offloaded.
- **Not a 250 Mbit/s cap.** That was the first number and it is misleading —
single-stream TCP. Four parallel streams doubled it. Quote the aggregate.
- **Not the interface.** wan1: `rxe=0 txe=0 rxd=0 txd=0`, no collisions.
## Most likely cause
Both tunnels negotiate **`aes256-sha1`** in phase 1 *and* phase 2 (dhgrp 14,
IKEv2). AES-CBC + SHA1 is a two-pass operation; FortiGate NPUs are markedly
faster on **AES-GCM**, which combines encryption and authentication in one
pass. The datasheet IPsec headline for an 80F assumes GCM with large packets,
not CBC+SHA1 at the 1438-byte tunnel MTU this link negotiates. The ~4x
shortfall is consistent with that.
## Not executed
Changing the proposal is a **production-edge change requiring a matching
change at the far end** (NH3 UDM and the ESH UDM), and each tunnel drops while
it renegotiates. Left for the operator. See the index entry for authorization
state.
## Immediate mitigation, no config change
Per-flow is the weak axis: a single stream over Site Magic gets 557 Mbit/s, a
single stream through IPsec gets 227. **Anything moving bulk data across the
Anaheim link should parallelise** — that alone roughly doubles throughput
today.
## Practical consequence already observed
`/mnt/smithy` mounted on ana-ml2 reads at 24.7 MB/s sequential vs 98.3 MB/s
from nh3-dev (same file, same mount) — that gap *is* this tunnel, not NFS and
not the NAS. See [[2026-08-23-smithy-mount-ana-ml2]].
## Access note
ana-gw is a FortiGate-80F, FortiOS 7.2.10, at 10.250.0.1. `sshpass` is absent
on nh3-dev; connect with paramiko via `uv run --with paramiko`. Password is
vaulted at `fortigate/ana-gw-infra-ops-password`. **`diagnose vpn tunnel list`
prints live ESP session keys** — never paste its output into althing, a
booth, or a commit.
---
## CORRECTION (2026-08-23, later session): the cutover was attempted and the remedy does not exist
The operator authorised the AES-GCM cutover, NH3 side first. It cannot be done,
and the measurements taken while trying show there is very little left to win.
### AES-GCM is unavailable on the far end — not a naming problem
The NH3 edge is a **UDM Pro SE** terminating `pfi-nh3-ana` (networkconf
`_id 697d64414c85dd2b6669b00a`, `ifname vti64`). Its UniFi API **validates** the
crypto enum and rejected every GCM spelling tried — `aes256gcm`, `aes256gcm128`,
`aes256gcm16`, `aes-256-gcm`, `aes256-gcm`, `aes256gcm12`, `gcm`, `aes128gcm128`
— all `HTTP 400 api.err.InvalidPayload`, nothing applied.
**The control that makes this conclusive:** the *identical* request body with
`ipsec_esp_encryption: "aes256"` returns `HTTP 200 rc:ok`. So the 400s are the
enum rejecting the value, not a malformed body. Corroborating: **zero
case-insensitive `gcm` matches across 7.3 MB of UniFi OS UI bundles.**
Accepted enum (probed): `aes128`, `aes192`, `aes256`, `3des` → 200; `des`,
`chacha20poly1305` → 400. There is no AEAD option. Both Anaheim tunnels land on
UniFi far ends, so this blocks the ESH tunnel too.
The FortiGate side **was** widened and is GCM-capable: phase2 `pfi-ana-nh3` now
reads `set proposal aes256-sha1 aes256gcm`. Left in place deliberately — it is
functionally identical while the peer only offers CBC, and reverting it would
cost another SA renegotiation for a cosmetic gain. Phase 1 was never touched;
IKE protects the control channel only and has no bearing on data throughput.
### The numbers that retire this as a problem
Measured NH3→ANA through the tunnel, and NH3→ESH over Site Magic (WireGuard) on
the same UDM and the same uplink, with the same stdlib TCP probe:
| streams | IPsec NH3→ANA | WireGuard NH3→ESH |
|---|---|---|
| 1 | 245 Mbit/s | 557 Mbit/s |
| 4 | 471 Mbit/s | 767 Mbit/s |
| 8 | **692 Mbit/s** | **795 Mbit/s** |
**NH3's WAN is a 1 Gbps link** (`uplink.speed = 1000`, port capable of 10G) —
that, not Anaheim's 2 Gbps, is the ceiling for anything crossing this tunnel.
So the tunnel does **~69% of the achievable uplink** at 8 streams, and the
IPsec-vs-WireGuard gap collapses from 2.3× at one stream to **15% at eight**.
Re-architecting the transport (site-to-site WireGuard via `ana-wg`, since
FortiOS has no WireGuard) would chase that last 15%. Not worth it.
### What the constraint actually is
A **per-stream** limit (~245 Mbit/s), not an aggregate crypto ceiling. Both
endpoints are idle at load — FortiGate CPU 100% idle with `npu_flag=03`
(offloaded both directions), UDM CPU ~7% with load1 moving 0.70 → 1.55. The
shape is per-SA/per-flow serialisation, and WireGuard shows the same shape from
a higher floor (557 → 795 is only 1.43× scaling).
### Actionable consequence
Anything moving bulk data across this link should **parallelise** — 245 → 692
Mbit/s, a 2.8× win with no config change. For single-stream workloads that
cannot be parallelised at the application layer, **NFS `nconnect=N` is the
lever**: it opens N TCP connections per mount, converting a single-stream
workload into a parallel one. The `/mnt/smithy` mount on ana-ml2 reading at
24.7 MB/s (~200 Mbit/s, i.e. exactly the single-stream ceiling) is the live
example — remounting with `nconnect=8` is the obvious test.
### Foot-gun recorded
Probing the enum by PUTting candidate values **applies the accepted ones**. A
probe loop here timed out with `3des` briefly live on the NH3 side, which the
FortiGate would not accept — a short tunnel outage until `aes256` was restored
(~1 minute, confirmed by the SA counters resetting). If you enumerate a UniFi
config enum this way, restore the known-good value after **every** 200, not at
the end of the loop. Post-change verification: the UDM object was diffed
field-by-field against its pre-change snapshot and is **byte-identical**.
---
## FOLLOW-UP (2026-08-23): what the per-stream limit actually is
The correction above called the constraint "per-SA/per-flow serialisation".
That was a hand-wave. Measured properly, it is a **hard per-flow rate cap of
~230–245 Mbit/s with a very deep buffer in front of it** — not a tuning
problem, not loss, not window size.
### The evidence: pin the send buffer and sweep it
Single stream NH3 → ana-docker, `SO_SNDBUF` pinned, `ss -ti` sampled in flight:
| in-flight cap | throughput | RTT in flight | minRTT | retrans |
|---|---|---|---|---|
| 256 KB | 224 Mbit/s | 7.8 ms | 5.3 ms | 0 |
| 416 KB | 225 Mbit/s | 11.8 ms | 6.6 ms | 0 |
| 416 KB | 245–247 Mbit/s | 12.0 ms | 5.6 ms | 0 |
| ~3.3 MB (autotuned) | 245 Mbit/s | **107 ms** | 5.5 ms | 0 |
**Throughput is flat across a 13× range of in-flight data while RTT scales with
it.** That is the signature of a fixed service rate with a standing queue: the
window controls only how much queue you build, never how fast you go. Had this
been window-limited, throughput would have risen with the buffer. Had it been
congestion, there would be retransmits — there are essentially none
(`retrans:0`, 0% ping loss).
So `net.ipv4.tcp_*` tuning, window scaling and congestion-control choice are all
**red herrings here**. Do not go there.
### Bufferbloat: one bulk stream wrecks latency for everything else
Measured on the same tunnel, ping to ana-docker:
- idle: **6.9 ms** avg
- during a **single** bulk TCP stream: **102 ms** avg, 136 ms max, 0% loss
**15× latency inflation from one transfer.** This is the operationally
important finding — any interactive traffic sharing the Anaheim link (ssh,
RDP, althing, VoIP) degrades badly whenever anything moves bulk data, and it
takes only one stream to do it. Parallelising transfers makes throughput
better and this *worse*. If it starts biting, the fix is an AQM/shaper on the
tunnel (or rate-limiting bulk jobs), not more buffer.
### Where the cap lives — strong inference, not proof
Three paths, and the FortiGate is the only variable:
| path | single-stream |
|---|---|
| FortiGate ↔ NH3 UDM (IPsec) | 245 Mbit/s |
| FortiGate ↔ ESH UDM (IPsec) | 249–265 Mbit/s |
| NH3 UDM ↔ ESH UDM (WireGuard, **no FortiGate**) | 557 Mbit/s |
Present in both slow paths, absent from the fast one. Aggregate over the same
SA reaches 692 Mbit/s, so it cannot be the SA or the crypto engine as a whole —
many flows spread out fine, one flow does not.
The mechanism that fits is **FortiGate NPU IPsec offload being per-session**:
each firewall session is bound to one crypto engine, so a single TCP flow is
capped at one engine's rate while many sessions spread across engines. **This
is inference from the throughput shape, not something confirmed on the box** —
`diagnose sys session list` was not captured for a TCP flow (the filter caught
only traceroute UDP probes). A single-stream control through ana-gw *without*
IPsec returned 290 Mbit/s to Hetzner Ashburn, but at ~60 ms RTT that is
window-limited and does not discriminate. **If this matters, the clean test is
a non-IPsec single stream between two Anaheim VLANs at low RTT.**
**Relevant to the FortiGate cutover decision:** if the per-flow cap is the
FortiGate's IPsec path, replacing the box plausibly lifts single-stream
throughput toward the WireGuard figure. That is a point in favour of the
cutover, and it is cheap to verify afterwards by re-running the sweep.
---
## FOLLOW-UP 2 (2026-08-23): it is NOT a capacity problem, and it IS specific to IPsec
Operator asked directly whether the 80F "can't handle the traffic". It can.
Two new measurements settle the shape of this, and correct an overstatement in
FOLLOW-UP 1 (which pointed at the FortiGate on evidence that was confounded —
every slow path was *both* IPsec *and* FortiGate, so protocol and box could not
be separated by that argument).
### The 80F routes a single flow at line rate when IPsec is not involved
`ana-ml2 → pfi-pve`, inter-VLAN **through** ana-gw (traceroute hop 1 =
`10.250.50.1`), 0.36 ms RTT, no tunnel:
| streams | throughput |
|---|---|
| 1 | **940.2 Mbit/s** |
| 8 | 939.3 Mbit/s |
Single stream saturates 1 GbE. So the box does **not** cap single sessions in
general, and there is no per-session ceiling in its plain forwarding path. The
~250 Mbit/s per-flow cap is **specific to the IPsec datapath**.
### Both IPsec tunnels converge on the same numbers despite different far ends
Measured today with the same probe:
| tunnel | far-end gateway | RTT | 1 stream | 8 streams |
|---|---|---|---|---|
| NH3 ↔ ANA | UDM Pro **SE** | 6.7 ms | 245 Mbit/s | 692 Mbit/s |
| ESH ↔ ANA | UDM Pro **Max** | 3.9 ms | **268 Mbit/s** | **715 Mbit/s** |
Different gateway hardware, different sites, different uplinks, and RTT
differing by 1.7× — yet single-stream differs by only 9%. **If this were
window-limited the 3.9 ms path would be ~1.7× faster.** It is not, which is
independent confirmation of a rate cap rather than a BDP effect.
### Capacity summary — the box has headroom it will not give one flow
- plain routing, 1 stream: **940 Mbit/s** (line rate)
- plain routing to internet, 8 streams: **2,153 Mbit/s**
- IPsec, 8 streams: **692–715 Mbit/s**
- IPsec, 1 stream: **245–268 Mbit/s**
- CPU **100% idle** throughout; IPsec NPU-offloaded (`npu_flag=03`)
Within a single SA, 8 sessions get ~2.9× what 1 session gets, so the datapath
distributes work **by inner session** — consistent with IPsec offload binding a
session to one crypto engine.
### What is still NOT separated
Whether the cap belongs to **the 80F's IPsec offload** or to **UniFi's IPsec
implementation**. Both tunnels have a UDM at the far end, and both UDMs run the
same UniFi firmware, so identical caps are explainable either way. The Pro Max
being only 9% faster than the Pro SE argues against the UniFi side (a beefier
CPU should show more), but that is suggestive, not conclusive.
**The test that closes it:** an IPsec tunnel whose endpoints do not include the
80F — e.g. a temporary UDM↔UDM IPsec tunnel between NH3 and ESH, measured
single-stream. If it also caps ~250, the FortiGate is exonerated and replacing
it buys nothing on this axis. If it runs near the 557 Mbit/s that UDM↔UDM
WireGuard achieves, the 80F is the limiter. **Bears directly on the pending
FortiGate cutover** — worth running before that decision, not after.
---
## FOLLOW-UP 3 (2026-08-23): WireGuard over the same internet path does 767 Mbit/s on ONE stream
Operator asked for a WireGuard test from `ana-wg` to NH3 over the public
internet. It is the test that separates the *path* from the *crypto*, and the
answer is unambiguous. **It also overturns FOLLOW-UP 1's "re-architecting the
transport is not worth it" — that conclusion compared 8-stream numbers and was
wrong for single-stream workloads.**
### Setup (fully torn down afterwards)
`ana-wg` (10.250.50.252, Debian 12 LXC, 4 cores) already has an
internet-reachable WireGuard endpoint: wg0 on **UDP 31337**, published by
FortiGate VIP `wg-to-ana-wg` (extip **38.120.12.42** → 10.250.50.252:31337,
policy 46, service `WireGuard-LEET`). **No FortiGate change was needed.** A
temporary `wgt0` was created on nh3-dev (10.30.10.200/32) as a fourth peer on
wg0, measured, then removed — ana-wg is back to its original 3 peers and the
keys were shredded. `wireguard-tools` was installed on nh3-dev and **left in
place** (benign, and wanted if this becomes permanent).
In this topology **neither gateway does crypto**: the FortiGate and the NH3 UDM
only NAT/forward UDP, and Linux does WireGuard at both ends.
### The full comparison
| path | crypto performed by | 1 stream | 8 streams |
|---|---|---|---|
| IPsec NH3↔ANA | FortiGate + UDM | 245 Mbit/s | 692 Mbit/s |
| IPsec ESH↔ANA | FortiGate + UDM | 268 Mbit/s | 715 Mbit/s |
| **WireGuard NH3→ana-wg** (same internet path) | **Linux + Linux** | **767 Mbit/s** | 763 Mbit/s |
| WireGuard NH3↔ESH (Site Magic) | UDM + UDM | 557 Mbit/s | 795 Mbit/s |
| plain routing through the 80F (inter-VLAN) | none | 940 Mbit/s | 939 Mbit/s |
**One stream equals eight streams over Linux WireGuard (767 ≈ 763).** There is
no per-flow penalty at all, and a single flow already saturates the path. So
the ~245 Mbit/s per-flow cap is **not** the ISP, not the circuit, not the NH3
uplink and not the physical path — all of which sustain 767 on one flow.
Per-flow penalty ranks by implementation:
- **Linux WireGuard — none** (767 → 763, flat)
- **UDM WireGuard — mild**, ~1.4× (557 → 795)
- **IPsec on this pair — severe**, ~2.8× (245 → 692)
### Latency under load — the same story
| path | idle | during ONE bulk stream |
|---|---|---|
| IPsec NH3↔ANA | 6.9 ms | **102 ms** avg, 136 ms max |
| WireGuard NH3→ana-wg | 6.2 ms | **12.7 ms** avg, 23 ms max |
WireGuard carries **3.1× the single-stream throughput with 8× less latency
inflation** on the same wire.
### Attribution — still not fully separated, and it no longer matters much
Both IPsec measurements have a FortiGate *and* a UDM doing IPsec, so this still
does not isolate which one imposes the 2.8× penalty. Closing that would need
Linux↔Linux IPsec or UDM↔UDM IPsec on the same path. **But the practical
decision no longer depends on the answer**, because the fix is the same either
way and it is already demonstrated.
### Recommendation (supersedes FOLLOW-UP 1)
A **WireGuard site-to-site between NH3 and Anaheim, terminated on `ana-wg`**, is
worth real consideration: 3.1× single-stream, flat scaling, far better latency
under load, and it reuses infrastructure that already exists and is already
internet-reachable. It is also the architecture already proven for NH3↔ESH.
Open questions before committing: routing/failover if ana-wg (an LXC) is down,
whether it replaces or parallels the IPsec tunnel, and firewall policy for the
new transit. ana-wg CPU was only ~40% busy across 4 cores at 767 Mbit/s, so it
has headroom.
**AND: `nconnect=8` on /mnt/smithy remains worth doing regardless** — it is the
same lever (turn one flow into many) and brokkr-smithy-dev has given standing
approval to apply it once the FortiGate work settles, with no need to ask again.
---
## RESOLVED (2026-08-23): it is the UDM's software AES-CBC. The FortiGate is exonerated.
Operator's theory — the UDM does IPsec in software with no crypto offload, so
the cost of the cipher itself is the limit — is **correct**, and it is now
demonstrated rather than inferred. He also correctly pointed out that
UDM↔UDM Site Magic is **WireGuard, not IPsec**, so that row never said anything
about UniFi's IPsec performance. It didn't, and I had leaned on it.
### The controlled experiment: vary cipher cost, hold everything else
AES-128 is 10 rounds, AES-256 is 14. If software crypto is the binding
constraint, throughput must rise when the cipher gets cheaper. If the limit
were the FortiGate's NPU, it would not move at all — hardware crypto is not
cipher-cost-sensitive in that range. Run A/B/A, single stream, 25–60 s each:
| condition | ESP cipher | single-stream | UDM CPU |
|---|---|---|---|
| A | aes256-cbc + sha1 | 232.3 Mbit/s | 35.4% |
| B | **aes128**-cbc + sha1 | **281.8**, 274.9 Mbit/s | 35.5% |
| A again | aes256-cbc + sha1 | 244.9, 242.5 Mbit/s | — |
**~1.16–1.20× faster on the cheaper cipher at identical CPU.** Same bytes of
CPU work, more payload through it. That is the signature of CPU-bound software
crypto, and it rules out the FortiGate's NPU as the limiter.
### Correcting two of my own earlier claims
1. **"UDM CPU is only ~7%, so it isn't CPU-bound" was WRONG — a sampling
artifact.** UniFi's `system-stats.cpu` refreshes on the device report
interval; 4-second sample windows were reading stale values. Under a
sustained 60 s single-stream load it reads **35.4%**, with load1 rising
0.60 → 1.17. On a 4-core UDM Pro SE that is ≈1.4 cores — one core saturated
on crypto plus overhead. **Always drive load for ≥60 s before trusting a
UniFi CPU figure.**
2. **The "FortiGate per-session NPU offload" hypothesis is REFUTED**, not merely
unproven. It predicts no change from a cipher swap; a 20% change was measured.
### Why the numbers all line up now
- **1 stream = 1 core of UDM crypto** → ~240 Mbit/s on AES-256-CBC.
- **8 streams = ~3 usable cores** → ~692 Mbit/s, ≈2.9× the single-stream figure
on a 4-core box. Aggregate is noisy (492–692 across repeats on a live link)
and is *not* cipher-sensitive, consistent with it being bounded by the path/
uplink rather than crypto once several cores are engaged.
- **AES-CBC is the specific villain: it is serial.** Each block depends on the
previous one, so the ARM AES instructions cannot pipeline across blocks. GCM
(CTR-based) and ChaCha20-Poly1305 both parallelise freely. That is why the
same UDM does 557 Mbit/s single-stream on WireGuard and only 240 on IPsec.
- **This retroactively vindicates the GCM cutover as the right idea aimed at the
right box** — GCM would have removed the serial dependency on the constrained
end. UniFi simply does not offer it, which is what made it impossible.
### Options this opens
- **AES-128 instead of AES-256: ~16–20% for free**, no topology change, one API
call per end. 128-bit is not the weak link here (SHA1 integrity is more
dated, and unchanged either way). Operator's call — **not adopted**, restored
to aes256.
- **WireGuard site-to-site via ana-wg: 767 Mbit/s single-stream** (3.1×), and it
sidesteps the UDM's IPsec datapath entirely. Still the biggest win available.
- Replacing the FortiGate **will not help this** — it was never the constraint.
Worth knowing before the cutover.
### State left behind
UDM network object verified **byte-identical** to its pre-test snapshot
(aes256/sha1). Tunnel up, selectors 1/1. FortiGate phase2 `pfi-ana-nh3` is
left as `aes256-sha1 aes256gcm aes128-sha1` — a permissive superset; the peer
offers only aes256 so the extra entries are inert, but **narrowing it back to
`aes256-sha1` is one line** if the looser list is unwanted.
---
## FOLLOW-UP 4 (2026-08-23): a downstream WireGuard terminator costs nothing to forward through
Operator's point: FortiOS has no WireGuard, so a WireGuard site-to-site must
terminate on a box *behind* the edge. Correct — and `ana-wg` (LXC, CT 113 on
pfi-pve, 10.250.50.252) already is that box.
**This closes a gap in FOLLOW-UP 3.** That 767 Mbit/s figure was measured with
traffic terminating *on* ana-wg. Real traffic must be forwarded onward to other
Anaheim hosts, which was never measured. Now it is:
| topology | 1 stream | 8 streams |
|---|---|---|
| IPsec, FortiGate ↔ UDM (today) | 245 Mbit/s | 692 Mbit/s |
| WG terminating **on** ana-wg | 767 Mbit/s | 763 Mbit/s |
| **WG transit: nh3 → wg → ana-wg → forward → ana-docker** | **763.8 Mbit/s** | **790.4 Mbit/s** |
**Forwarding through the LXC is free** (763.8 vs 767). The downstream-VM
architecture delivers the full 3.1× single-stream for real transit traffic, not
just for traffic landing on the tunnel box.
ana-wg while forwarding 764 Mbit/s: **~22% busy across 4 cores** (77.8% idle),
so roughly 0.9 cores. Note `/proc/loadavg` inside this LXC reports the *host's*
load, not the container's — do not read it as ana-wg's own. For contrast the
UDM burns 35.4% of its 4 cores to move 240 Mbit/s, so ana-wg has ample headroom.
### Design consequences of terminating downstream — the parts that need decisions
1. **Anaheim hosts must route to ana-wg, not to the FortiGate.** The 763.8
figure was obtained with an explicit `10.30.10.200/32 via 10.250.50.252`
route on ana-docker. Without that, a host sends 10.100.0.0/16 to its default
gateway (ana-gw), which routes it back out the *same* interface to ana-wg — a
LAN hairpin crossing the FortiGate twice. **The hairpin variant was NOT
measured.** Options: DHCP option 121 pushing the route fleet-wide, a dedicated
transit VLAN for ana-wg, or accept the hairpin.
2. **New single point of failure.** Today site-to-site dies only when the edge
dies, which is total anyway. A downstream terminator fails independently.
Mitigation: keep the IPsec tunnel configured as a higher-metric fallback
route so it takes over when ana-wg is down.
3. **ana-wg is an LXC on pfi-pve**, so its ~0.9 cores and NIC traffic land on the
hypervisor shared with the rest of the Anaheim VMs.
4. **The NH3 end needs a terminator too**, and there are two shapes:
- **Linux VM at NH3** (nh3-dev or a dedicated VM on nh3-pve) — this is what
was measured: **764 Mbit/s**.
- **NH3 UDM's existing WireGuard server** (`PFI-NH3-WG`, wireguard-server on
UDP 31337) accepting ana-wg as a peer — plausible but **untested**, and
UniFi's WireGuard shows a per-flow penalty (557 Mbit/s single-stream on
Site Magic), so expect ~557 rather than 764. Still 2.3× today.
### Standing recommendation
Worth doing, but it is **a project, not a config tweak** — routing, failover and
policy all need deciding. The cheap wins remain available meanwhile and are
independent: `nconnect=8` on NFS mounts (approved by brokkr-smithy-dev, pending
the FortiGate work settling) and AES-128 for ~20%.
---
## LANDED (2026-08-23): AES-128 on both tunnels; FortiGate public admin closed
Operator directed: adopt AES-128 on **both** Anaheim tunnels, make-before-break,
then close the FortiGate's WAN and SSH admin surfaces. All done and verified.
**Context that retires the WireGuard-in-a-VM design work:** the FortiGate is
being **replaced by OPNsense on a Dell R420**, which gives **WireGuard on the
edge device itself**. The downstream-terminator architecture (FOLLOW-UP 4) is
therefore moot — do not scope it. This also **un-parks the OPNsense migration**,
which auto-memory recorded as PARKED pending "hardware acquisition"; the R420
is that trigger.
### What changed
Make-before-break on the FortiGate first, so neither tunnel dropped waiting on
a far end:
| phase2 | proposal now |
|---|---|
| `pfi-ana-nh3` | `aes256-sha1 aes256gcm aes128-sha1` |
| `ana-eshudm-dyn` | `aes256-sha1 aes128-sha1` |
Then each UDM flipped to `ipsec_esp_encryption: aes128`:
| tunnel | UDM object | before | after |
|---|---|---|---|
| NH3 ↔ ANA | `pfi-nh3-ana` `697d64414c85dd2b6669b00a` @ 10.100.0.1 | 245 Mbit/s | **269.7** |
| ESH ↔ ANA | `esh-ana` `697723b9b9d4266dddf2bcc7` @ 10.0.0.1 | 268 Mbit/s | **304.3** |
Single-stream gain ~10–13% here, against 16–20% in the earlier controlled A/B —
the difference is live-link variance, not a different result. Both UDM objects
were diffed field-by-field against pre-change snapshots: **the only field that
moved on either is `ipsec_esp_encryption`.**
The FortiGate proposal lists were deliberately **left permissive** (still
accepting aes256). The peers offer only aes128 so the extra entries are inert,
and keeping them means a UDM reverting does not strand the tunnel. Narrowing to
`aes128-sha1` alone is a one-liner if the looser list is unwanted.
### Admin surfaces closed
`wan1 allowaccess` → **`ping`** (https + ssh removed) and `infra-ops` trusthost
→ **10.0.0.0/8 only** (the 8 wide-open ranges unset). Verified 443 and 22 closed
from both NH3 and ESH; management over the tunnel at 10.250.0.1 still works.
**Sequencing that matters: the close was executed over the TUNNEL path, not over
WAN** — removing `ssh` from allowaccess while connected over WAN kills the
session mid-command.
**Consequence to hold in mind: ana-gw now has no out-of-band management path.**
If both tunnels drop it is console-only until someone is on site.
### Gotcha: the two UDM vault items have DIFFERENT shapes
- `unifi/pfi-udmse-api-key` → a **bare 32-char key**. `secret get` output is the key.
- `unifi/esh-udmpm-api-key` → a **19-line documentation note** with the key on a
`key:` line. `secret get` piped straight into a header yields a 1396-byte
value and the UDM answers **`400 Bad Request` from nginx**. Extract with
`grep '^key:' | awk '{print $2}'`.
**The ESH key's first-ever confirmed WRITE happened here** (auto-memory recorded
it as read-verified only): a control PUT of the unchanged object returned
`rc:ok`, then the real change did too. That key has a full read+write admin role.
---
## CORRECTION (2026-08-23): port 80 on the WAN IP is the FortiOS ACME listener
The claim in the previous section that `.42:80` was an **ISP transparent proxy**
was **WRONG**, and so was the earlier warning that ACME renewal would fail with
port 80 absent from `allowaccess`. Operator pushed back asking where the port-80
map terminated. It terminates **on the FortiGate itself**.
**What it is:** the FortiOS **ACME HTTP-01 challenge listener**. `config system
acme` has `set interface "wan1"`, and FortiOS opens port 80 on that interface to
answer Let's Encrypt challenges **independently of `allowaccess`** — `wan1
allowaccess` reads `ping` only and the port is still open. Every non-challenge
request returns a fixed 403 whose body is literally:
```
<!DOCTYPE html><html><head><title>ACME Access Only</title></head><body>ACME Access Only</body></html>
```
**Not a DNAT.** The full VIP table has 14 entries; only two land on `.42` —
`Kokoro-In` (:8880 → 10.250.50.51) and `wg-to-ana-wg` (:31337 → 10.250.50.252).
~~Worth noting separately: four VIPs are all-port static NAT~~ — **that claim was
WRONG, see the correction below.** All fourteen VIPs are scoped.
### The methodology error that produced the wrong answer — worth not repeating
The sniffer filter used was `dst host 38.120.12.42 and tcp port 80`. **`dst host`
matches only inbound packets**, so outbound SYN-ACKs were excluded *by
construction*; concluding "the box sends no SYN-ACK" from that capture was
unsound. Re-run with the bidirectional `host 38.120.12.42 and tcp port 80` it
immediately shows `wan1 out 38.120.12.42.80 -> <scanner>: syn ack`.
**Rule: when testing whether a box *answers*, the sniffer filter must be
bidirectional. `dst host` silently answers a different question.**
### Consequences
- **ACME renewal will work** with `allowaccess ping`. The earlier "add `http`
back or the cert expires" warning is retracted — FortiOS opens the challenge
port itself. Cert valid to 2026-10-27, renewal attempt ~2026-09-27.
- **It is not an admin surface** — static 403, no auth, no GUI.
- Its practical value is now low: WAN admin is closed, so the cert only serves
the internal GUI at 10.250.0.1, where the name would not match anyway. Killing
it (`config system acme` → unset interface) would close the last WAN listener
at the cost of cert renewal. Operator's call; **not done**.
---
## CLOSED OUT (2026-08-23): ACME disabled; and the "all-port VIP" alarm was FALSE
### ACME disabled — the WAN IP now exposes nothing
`config system acme / unset interface` (the account object is left in place;
with no interface bound there is no listener). Verified:
- **External scan of 38.120.12.42 across 55 ports: no open TCP ports at all.**
- Internal GUI at 10.250.0.1 still answers **200**, SSH still works.
- `admin-server-cert` is still `ana-fw.pfi` — the existing cert is untouched and
serves the internal GUI until **2026-10-27**; it simply will not auto-renew.
Reverse with `config system acme / set interface "wan1"`.
### RETRACTION: the four VIPs are NOT all-port
A previous section claimed `Rustdesk`, `https-to-tacticalrmm`, `web-to-webhost`
and `web-to-sfcontainer` were unrestricted all-port static NATs. **They are not.**
A FortiOS VIP can be scoped **two different ways** and the parser used only
checked one:
1. `set portforward enable` + `set extport <n>` — a single mapped port, **or**
2. `set service "<svc>"` on the VIP object — constrains the VIP to that service.
All four use form 2. The custom services are narrow: `Rustdesk` = TCP
21115–21119 + UDP 21116 (the standard RustDesk range), `ssh-mapped-2223` = TCP
2223 only. **Every one of the 14 VIPs is scoped; none is unrestricted.**
**Lesson: absence of `portforward` does NOT mean all-port on a FortiOS VIP —
check `service` too.** Better still, do what settled it here: scan from outside
rather than reading config.
### Ground-truth public exposure (external TCP scan, post-change)
| IP | open | maps to |
|---|---|---|
| 38.120.12.41 | *nothing* | — |
| **38.120.12.42** | ***nothing*** | the FortiGate itself — fully closed |
| 38.120.12.43 | 80, 443 | sf-ana-container 10.250.150.100 (SureFire tenant) |
| 38.120.12.44 | 22, 80, 443, 8025, 21115–21119 | gitea (→222), traefik, mailrise, RustDesk |
| 38.120.12.45 | 80, 443, 2223 | pfi-ana-webhost 10.250.50.52 (2223→22) |
| 38.120.12.46 | 443 | pfi-tacticalrmm 10.250.50.57 |
Configured-but-closed: 8443 (mattermost-calls), 8444 (webdav-nas), 8880
(Kokoro-In) — VIPs exist, nothing listening behind them. Worth a tidy-up during
the OPNsense translation but not exposure.
@@ -1,63 +0,0 @@
# [2026-08-23] hrafn adopted; its CI deploy reported green while deploying nothing
`hrafn` — genuine-Chromium browser-fetch behind a REST API, for bot-gated sites
(Reddit first). Built by nevermore-claude on ana-docker, handed to infra-ops for
uptime ownership. Internal-only on `traefik-net`, no host port; consumers reach
`http://hrafn:8080`. Canonical at `stacks/hrafn/`.
## Intake found a live credential exposure
`/opt/docker/compose/hrafn/.env` was mode **0644 with a live 57-char bearer token**
— verified as real exposure by reading it as `nobody` on a box with four
interactive accounts. Tightened to 0600. That triggered the wider sweep (see
[[2026-08-23-ana-docker-env-perms-sweep]]).
## The CI defect — the one worth remembering
I authored the deploy (elway playbook + gitea workflow) to replace a hand-rsync,
tagging the image with the commit SHA for provenance. nevermore-claude later found
v1.0.0 deploying "green" while the host still served 0.1.0.
**Root cause was mine and nastier than either hypothesis.** The staging dir was
`$compose_dir/.stage` — **inside** the rsync target. So
`rsync -a --delete $compose_dir/.stage/ $compose_dir/` deleted `.stage` from the
destination (absent from the source listing) **during** the transfer, destroying
its own source mid-copy. Reproduced exactly:
```
before: app.py="OLD" leftover.txt .stage/app.py="NEW"
after: app.py="OLD" leftover.txt GONE, .stage GONE
```
Deletion succeeded, the copy silently did not, rsync exited 0. So the directory
*looked* converged while host source stayed frozen at the first manual rsync —
and because the build's `COPY` inputs never changed, Docker full-cache-hit and
every SHA tag aliased one image. **The provenance the tagging existed to provide
was false for the pipeline's entire life.**
**The real failure is the verification.** The verify steps asserted the marker,
container health, and a 200 from `/readyz` — all of which pass against a
completely frozen host. None measured *content*. A deploy that reports success
without asserting the bytes changed is verifying an **uptime**, not a deploy.
## Fixes
- stage at `/tmp/hrafn-deploy-stage`, outside the target
- CI computes `context_sha256` over the shipped file list; the playbook recomputes
it **on the host after the converge** and fails on mismatch
- compare the running container's `src/**/*.py` against the host's, so a SHA tag
cannot name layers the image lacks
- **compare `*.py` only** — `pip install .` generates `src/*.egg-info/*` inside the
image and `__pycache__` appears at runtime, so a naive `find src -type f` compare
false-fails on every healthy deploy. Verified against a known-good container
before shipping (12 host files, 18 in container, 0 content differences).
- declined `--no-cache`: a cache hit is *correct* when the context is genuinely
unchanged; assert the property rather than brute-force it.
## Access
Operator granted claude-bot **write** on `vh/hrafn`, so infra-ops maintains the
pipeline it owns instead of routing patches through the repo holder. `vh/hrafn` is
canonical; `stacks/hrafn/ci/` is a verified mirror.
Commits `b6924de`, `b001d0c`, `11b9d18`, `b38c369`, `9642952`.
@@ -1,81 +0,0 @@
# [2026-08-23] selene seat retired after losing a head-to-head; 7 aliases share one seat
## Why selene went
Benchmarked against `gen` on selene's own job — 24 designed judge items with
checkable ground truth, pairwise + absolute modes, 3 repeats, run on **both** a
neutral JSON prompt and Selene's **native Atla template** (288 calls, free local).
```
neutral JSON selene 20/24 (83%) gen 23/24 (96%)
native Atla selene 21/24 (88%) gen 22/24 (92%)
```
gen won on both templates and **selene's best sat below gen's worst**. Selene was
given its own fine-tuned template as a fairness check before any recommendation;
it gained one point, not three.
**Decisive defect: selene cannot emit "tie"** — 0/2 on both templates, forcing a
winner on every equivalent pair. For eval work that is the case that matters.
brokkr-smithy-dev independently corroborated from the other end with a **null
control** (an excerpt compared against ITSELF, where tie is definitional):
`chat-judge`(selene) TIE **27/60 = 45%**, gen **60/60 = 100%**; ground-truth
recovery on real-corpus ranking selene **47% — chance** vs gen 94%. My 83-vs-96
understated it: on a *ranking* task selene was a coin flip. Absolute scoring on
designed items is an easier task than ranking real text — the harness is a
**screen, not a verdict**, and its README says so.
Reclaimed **17.2 GiB** on ana-ml2 GPU1 (free 1,818 -> 19,450 MiB).
## The naming rule, restated the hard way
I proposed repointing `selene-1-mini-8b` at gen and was **correctly overruled**:
> never repoint a named model at a different model's endpoint — that is
> intentionally misleading
`chat-judge` is a **role** alias (ADR-0012: consumers bind the capability) and
moved to gen with a deterministic judge profile copied from `image-judge`.
`selene-1-mini-8b` is a **model** name and was removed outright — it now returns
`HTTP 400 Invalid model name`, verified. The discriminator: *does the string
promise a capability, or an identity?*
## The 7-way alias collision — the finding with the longest reach
```
chat-judge classifier gen image-judge
qwen-image-bench summarizer summarizer-large -> qwen3.8-27b-uncensored :8015
```
Also colliding: `gen-frontier`/`gen-frontier-reasoning`/`glm-5.2`/`glm-5.2-reasoning`;
`ext-tts`/`gpt-4o-mini-tts`/`tts-1`/`tts-1-hd`; `reranker`/`reranker-a3-bge-v2-m3`.
**Cross-checking a result against another alias measures nothing when they are the
same weights — agreement is an echo, not corroboration.** Documented at the head of
`model_list` in the live gateway config, because it belongs where people read it.
This caught a real defect within hours: brokkr's R47 premium-corpus gate was about
to run ~46,000 record-exposures against `gen` with `summarizer` shortlisted as an
independent second opinion. They pinned the backing model in the preregistration
and dropped the second-alias idea instead.
## Provenance seam (brokkr's pushback, adopted)
The gateway returns the **alias** in the response `model` field, not the backing
model — so a per-call guard catches a swap *during* a run and is blind to one
*between* runs. **Role alias for routing, concrete model for provenance.**
`GET :4000/model/info` with the shared key already exposes backing model +
api_base; resolve at run start AND end and void on mismatch.
## Artifacts
- Harness kept at `tools/judge-bench/` (`--models` REQUIRED — a stale default
would silently benchmark a retired seat).
- `stacks/selene/` keeps compose + a README explaining the retirement.
- Technique worth stealing, from brokkr: **a control constructed so the correct
answer is DEFINITIONAL rather than judged cannot inherit the designer's error.**
Item vs itself; response vs its own truncation; text vs its own clauses
permuted. Add those before adding more judged items.
Commits `ca3c984`, `b8a5355`.
@@ -1,70 +0,0 @@
# [2026-08-23] Worldtree b187 shipped; all three instances de-armed from a 69-day-stale `:latest`; Matrix homeserver re-plumbed
## b187 pre-stage (#405 phases 1+2)
The matrix bridge stopped embedding the engine and became an HTTP client of the
Conversation API, so `WORLDTREE_API_URL` became **boot-blocking** — absent from the
container env, the bridge exits by design. Demo's compose never passed it; the next
recreate would have crash-looped. Pre-staged on demo and personal (additive, backed
up, verified with `docker compose config`, nothing restarted).
**Key decision, and I got its scope wrong first.** I argued demo should stay keyless
(no homeserver -> no rooms -> no turns -> no 401s). Right about turns, **wrong about
scope**: the engine preflight authenticates at boot regardless of homeserver, so demo
booted permanently degraded. Corrected — key `341c1488` minted under worldtree-dev's
recorded authorization, vaulted, wired, three-hop hash-verified.
## The 69-day-stale `:latest` landmine
All three instances pinned `WORLDTREE_IMAGE=.../worldtree:latest` in `.env` while
running SHA-tagged images built that day. Local `:latest` = `b19afd71d7cc`, built
**2026-06-14**. So ANY `docker compose up` — anyone's, for any reason — silently
downgraded that service by 69 days. Same footgun as the 2026-06-15 outage.
Re-pinned all three to their running SHAs (Worldtree #410), verified by rendering
compose config rather than reading `.env`, containers untouched. Playbook at
`playbooks/repin-worldtree-image.yaml`.
**`worldtree-pinned` was the worst case:** the instance whose entire purpose is being
frozen was running a **dangling image with no repo tags**, kept alive only by the
running container. One `docker rm` from garbage collection. Tagged
`:446e5807bf43` first, then pinned.
The guard I wrote had two bugs the pinned case exposed: it compared the container's
`.Config.Image` **string** (only the tag it was CREATED from — pinned was created
from `:latest` back when that meant 446e5807), and it reported CHANGED
unconditionally. Now compares **image IDs** and skips when already correct.
## Matrix homeserver ownership
Operator ruled: **personal owns the Matrix bridge.** The appservice tokens were never
missing — both sat at length 64 in the vaulted dev `env.sh` while both deployed
instances had them at length **zero**. Someone wired four of six Matrix vars and
stopped. Wired them into personal, three-hop verified.
**The trap worth remembering:** Synapse's registration pointed at
`http://10.100.10.50:8009` — nh3-dev, a dead epoch, with transaction 2801 queued at
512s backoff. The natural fix (swap the IP) gives `10.250.50.152:8009` which is
**DEMO's** bridge, and Synapse can reach both — it would have connected, delivered,
and looked correct while routing the operator's live rooms to the demo instance.
**Personal's bridge is :8010.** `docker port` is ground truth.
Corrected the URL, restarted Synapse (healthy in 32s after 3.5 months up), verified
`GET /_matrix/app/v1/ping -> 200` from inside the Synapse container. worldtree-dev's
smoke passed first try: room created, mimir accepted the invite, a real engine turn
ran, mimir replied in persona voice. #408 closed.
## Open on worldtree-dev's side
- **#411** — personal's bridge logs `Debug sink init failed: Permission denied:
/app/sessions/debug_rooms.json`. It creates two debug rooms but cannot persist
their IDs, so **every restart mints a fresh pair on the live homeserver**. Room
litter that compounds silently. Needs a which-container-writes-what check on the
sessions volume before anyone chowns it.
- Bridge/engine agent-roster drift: 6 of the bridge's 9 configured agents are not
listed by the engine on either instance.
- Historical Domari pairwise verdicts from the selene era are coin-flip-grade
(see [[2026-08-23-selene-retired-alias-collision]]); worldtree-dev banked that so
no future arc leans on them without re-judging.
Commits `064181a`, `bb19a96`.
@@ -1,50 +0,0 @@
# [2026-08-24] ana-gw public admin surface closed to zero, ACME listener included
WAN admin was opened at the start of the session as a cutover contingency
("so I don't have to drive down there"), then closed again on operator
instruction once the AES-128 work landed. Net result: **the FortiGate's WAN
address now exposes no TCP port at all.**
## Final state
External scan of `38.120.12.42`, 55 ports: **nothing open**. Verified from two
sites. `wan1 allowaccess` = `ping`; `infra-ops` trusthost back to `10.0.0.0/8`.
**Consequence to hold: there is no out-of-band path to ana-gw.** If both tunnels
drop it is console-only. Re-open is two one-liners (allowaccess + trusthost) —
both are recorded in auto-memory `reference_fortigate_ana_gw_access`.
## Port 80 was the FortiOS ACME listener, and I got it wrong first
`38.120.12.42:80` answered a bare 403 (`ACME Access Only`, 101 bytes) with
`allowaccess` set to ping only. First diagnosis — "an ISP transparent proxy" —
was **wrong**, and the reason is worth keeping:
> The sniffer filter was `dst host 38.120.12.42 and tcp port 80`. **`dst host`
> matches inbound only**, so outbound SYN-ACKs were excluded *by construction*,
> and concluding "the box sends no SYN-ACK" from that capture was unsound.
Re-run bidirectionally (`host … and tcp port 80`) it immediately showed
`wan1 out 38.120.12.42.80 -> <scanner>: syn ack`. **Rule: to test whether a box
*answers*, the filter must be bidirectional.**
The listener is opened by `config system acme / set interface "wan1"` and
**bypasses `allowaccess` by design** — FortiOS needs port 80 for HTTP-01. It
was disabled (`config system acme / unset interface`); the LE cert (`ana-fw.pfi`,
valid to 2026-10-27) is untouched and simply stops renewing, which is fine
because WAN admin is closed and the box is being replaced.
## Retracted in the same pass: the "four all-port VIPs" alarm
Claimed four VIPs were unrestricted all-port static NAT. **False.** A FortiOS
VIP is scoped **two** ways — `portforward`+`extport`, *or* a `service` binding
on the VIP object — and only the first was checked. All 14 VIPs are scoped;
`Rustdesk` is TCP 21115–21119, `ssh-mapped-2223` is TCP 2223 only.
Ground-truth external scan of all six public IPs is recorded in
`reference_fortigate_ana_gw_access`. Configured-but-dead: `:8443`
(mattermost-calls), `:8444` (webdav-nas), `:8880` (Kokoro-In) — tidy-up
candidates for the OPNsense translation, not exposure.
**Lesson, twice in one session: measure from outside instead of parsing config.**
Both wrong answers came from a filter that answered a different question.
@@ -1,167 +0,0 @@
# `[2026-08-24]` char-rp seat: OOM root-cause, Gemma-4 MoE swap, and the abliterated trainee base
One evening, one thread with brokkr-smithy-dev, five commits: `850e0c3`,
`27155c0`, `f509668`+`24e8826`+`1bd90ea`+`3446367`+`8d6a939`, `14ff4a3`,
`019ccff`, `5415fd4`.
## 1. The seat was crash-looping, and the cause was NOT its config
`vllm-meromero-rp` reported up-but-unreachable, RestartCount climbing (13 by the
time it was examined, not the 4 first reported). Startup logs looked clean all
the way through weights, `torch.compile` and CUDA-graph capture, then:
torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 336.00 MiB.
GPU 0 has a total capacity of 94.97 GiB of which 195.19 MiB is free.
**⚠ `--gpu-memory-utilization` SIZES THE KV CACHE AND DOES NOT COVER CUDA
CONTEXT, GRAPHS OR NON-TORCH OVERHEAD.** gen is configured at 0.43 and actually
held 45.6 GiB. char-rp was at 0.51. The pair was committed to 0.94 of the card
with ~0.6 GiB of real headroom — it fit on the 21st and stopped fitting on the
24th.
Dropped char-rp to 0.47: ~4.8 GiB margin, KV 27.36 → 23.56 GiB, 430,825 →
371,023 tokens against a max-model-len of 262,144. **Cost nothing usable** — the
pool still holds 1.4x a full-length sequence; what is lost is concurrent long
requests, not context.
**⚠⚠ THE MISSING HALF, found later that evening: gen's footprint GROWS WITH
UPTIME.** Same container, same 0.43: **45.6 GiB after ~3 days up, 38.5 GiB
freshly restarted** — ~7 GiB apart. Nothing about char-rp changed between the
21st and the 24th; *gen crept up underneath it*. **Headroom arithmetic done
against a long-running gen is measuring a moving number.** Measure against a
freshly-restarted one.
## 2. `char-rp` and `char-rp-reasoning` are ONE seat, not two
Both LiteLLM routes point at `10.250.50.54:8016/v1` — `hosted_vllm/char-rp` and
`hosted_vllm/char-rp-thinking`. brokkr had reported 30/80 and 80/80 failure
rates as two failing services; it was one outage sampled twice. This also
*improved* a result of theirs: their CoT on/off battery had assumed both aliases
were the same weights under two chat templates, and the routing detail turned an
assumption into a verified fact.
(`vllm-charrp-reasoning-nvfp4`, the Heretic2 NVFP4+MTP container, has been
stopped for 12+ days and is unrelated — it is not what that alias resolves to.)
## 3. The seat swapped to the Gemma-4 26B-A4B MoE
Operator-directed straight-across replacement: same port, same
served-model-names, so no gateway route or consumer config moved. Rationale is
throughput under CoT — the user waits through the whole reasoning block before
the first visible token, and the MoE measures ~114 tok/s @32K against the dense
31B's ~40.7.
Serving copy is `RedHatAI/gemma-4-26B-A4B-it-NVFP4` (16 GB), chosen over the
other `-it` quants because it is compressed-tensors (`nvfp4-pack-quantized`) —
the same loader path the outgoing seat used. Smaller weights at the same 0.47
budget bought **1,724,110 KV tokens against the predecessor's 371,023**.
`meromero-charrp` is retained stopped in `created` state, labelled
`AI - Dormant`. Both stacks bind `:8016`, so rollback is **stop-then-start**.
## 4. ⚠ THE STALE-CHAT-TEMPLATE TRAP IS ENDEMIC, NOT A ONE-OFF
Verified by hash across every third-party Gemma-4 derivative pulled:
| build | lines | sha256 (normalised) |
|---|---|---|
| upstream `google/gemma-4-26B-A4B-it` | 390 | `6a1015c47ccfcfa6` |
| RedHatAI NVFP4 (served) | 389 | `6a1015c47ccfcfa6` — the only match |
| llmfan46 heretic | 365 | `0a52be69cda5ab8a` |
| TrevorJS abliterated | 266 | `58c66fdee4afa297` |
| jenerallee78 abliterated | 266 | `58c66fdee4afa297` |
| prithivMLmods NVFP4A16 | 266 | `58c66fdee4afa297` |
Three independent repos carrying the *identical* stale file means it propagated
through the ecosystem. Consequences differ by use and **both are silent**:
serving a mismatched template renders a different prompt; **training through
`base/chat_template.jinja` means training on a different prompt format than
production serves** — train/serve skew, no error, presents as a tuning failure.
The production compose now pins the template explicitly. It is a **no-op for the
served weights** (the A4 build ships that exact file) and permanently closes the
class. ⚠ If `GEMMA4_MODEL` ever points at a different checkpoint, the template
default must move with it.
## 5. A benchmark result was RETRACTED — below chance indicts the instrument
A battery appeared to show Gemma at **12% contradiction detection with CoT off
against gen's 81%**. An A16 activation-precision control was staged to test
whether the quant scheme owned it. Then the operator asked to see the individual
items, and the item was **ill-posed**: it presented two mutually contradicting
statements and asked for "*the* contradicting statement", but **contradiction is
symmetric**. The model consistently named the absolute claim — a defensible
reading the labelling scored wrong every time.
**⚠ THE TELL WAS IN PLAIN SIGHT: 12% ON A FIVE-OPTION TASK IS BELOW THE 20%
CHANCE FLOOR.** A below-chance score indicts the instrument before it indicts
the model, and neither side reacted to it. I spent the afternoon verifying repo
names, config fields, template hashes and tokenizer vocabs — every layer of
plumbing — and never asked whether the number itself was *possible*. **A
preflight can be thorough and still be aimed in the wrong direction.**
Retracted: "the model owns the contradiction deficit"; "domain tuning costs 43
points of contradiction detection" (on a sound instrument it **reverses**); all
pre-fix T2 numbers. Recorded as a dated superseded-claims table in
`stacks/gemma4-charrp/README.md` rather than a silent edit.
**What survived:** the A16 control result — activation precision is close to free
on this battery, every other task identical across W4A4 and W4A16 builds.
## 6. The abliterated trainee base — measured, not assumed
Operator directed a low-damage abliterated instruct build. "Low damage" was
treated as a measurable claim; the field spreads from KL 0.09 to 0.4118:
| build | method | KL | refusals |
|---|---|---|---|
| **llmfan46** (operator's pick) | Heretic v1.2.0 ARA | 0.1237 | 3/100 |
| TrevorJS | ARA-family | 0.09 | 1/100 effective, 5/686 cross-dataset |
| jenerallee78 | ARA 2-pass | 0.1299 | 7.7% StrongREJECT |
| huihui-ai | remove-refusals | none published | none published |
Fleet anchor: our own work found **Heretic at KL 0.12 preserved the MTP head at
83.7% acceptance**, so both staged builds sit inside an already-measured band.
huihui-ai rejected — no metrics, its card calls the method "a crude,
proof-of-concept implementation", it abliterates both thinking and non-thinking
modes, and its parameter count runs ~738M over upstream. Operator's independent
read matched ("huihui produces garbage").
**Abliteration isolated properly** (stock BF16 vs llmfan46 BF16, same precision,
same pinned template, same 192 items):
T2 contradiction 75% → 59% (−5 items)
T6 spatial 75% → 88% (+4 items)
core 90.0% → 89.4% (−0.6 pts)
**It MOVED capability rather than removing it** — five lost on contradiction,
four gained on spatial, nearly cancelling. Nobody predicted a gain. **llmfan46
stands**; no case for re-staging on TrevorJS over 0.6 points.
⚠ Read as ~5 and ~4 items at n=32, not as −15.6/+12.5 percent. ⚠ Says nothing
about quantization — the stock-NVFP4 T2 figure came from n=16 against n=32,
different item sets, n-confounded.
## 7. ⚠ The production compose hardcodes `--quantization compressed-tensors`
Pointing the char-rp stack at unquantized BF16 weights crash-loops immediately:
TypeError: CompressedTensorsConfig.__init__() missing 3 required
positional arguments: 'target_scheme_map', 'ignore', 'quant_format'
vLLM trying to read a quantization config out of a checkpoint that has none. 35
restarts before it was caught. Hence `stacks/gemma4-trainee-bench/` — a separate
ephemeral stack with no quantization flag, `restart: "no"` so a bench seat cannot
resurrect itself and block gen's restore, and no homepage labels so it leaves no
permanently-offline card. That detour is why a base swap is now ~5 minutes
instead of ~15.
## 8. BF16 cannot coexist with gen
48.07 GiB of BF16 weights plus gen's footprint exceeds the 94.97 GiB card before
a byte of KV cache. Every BF16 bench window means **gen is stopped**. Two such
windows were run and gen restored both times; the restore was triggered by
observing the seat's own throughput logs (a large prefill burst then zero
running/zero waiting) rather than waiting on a courtesy message.
Cross-links: [[2026-08-24-homepage-uniform-grid]]
@@ -1,77 +0,0 @@
# [2026-08-24] ESH DNS fixed at the IPv6 layer, and the naming scheme went live
Reported as "`scriberr.ana.internal` doesn't resolve on my Mac, and nslookup
shows an IPv6 DNS server." Operator's diagnosis was right; the fix took three
wrong turns worth recording.
## Root cause
`esh-userland` has IPv6 PD with RA at `pref high`, and the UDM advertises
**itself** as the resolver via RDNSS. macOS honours RDNSS and prefers it over
the DHCPv4-supplied resolver, so queries went to the UDM — which does not know
`.internal` — and returned NXDOMAIN. AdGuard was never consulted.
Two adjacent gaps found while there: `esh-userland`'s **secondary** v4 resolver
was `10.0.10.1` (the UDM itself), and `esh-server` had **DNS handout disabled
entirely**, so every host there got the UDM and could never resolve `.internal`
— esh-docker-vm was living proof.
## The three wrong turns
1. **`dhcpdv6_dns_auto=false` alone does nothing.** It is only honoured **when an
explicit server is supplied**. Setting it bare looks like a no-op and invites
the conclusion that the field is inert — which I drew, wrongly. Despite the
`dhcpdv6_` prefix it *does* drive the RA's RDNSS option on a SLAAC network.
2. **`wan_dns1` is NOT used by the UDM's LAN-facing forwarder.** Setting it to
AdGuard persists, reads back, and changes nothing. Proven with **fresh
uncached ad domains** — AdGuard blocklists answer `0.0.0.0`, the UDM returned
real IPs. Reverted.
3. **`force-provision` returns `rc:ok` and proves nothing** — consistent with the
known `cmd/devmgr` behaviour.
Every failed attempt returned `rc: ok`. **Verify by observed effect.** RAs were
probed with a stdlib raw-socket Router Solicitation parsing option type 25
(`rdisc6`/`tcpdump` were both absent; nothing was installed).
## What landed
RDNSS **redirected** rather than disabled — better than switching it off:
| VLAN | v4 | v6 RDNSS |
|---|---|---|
| `esh-userland` | 10.0.50.45 + 10.100.50.40 | `…:4411:b105:50:45` |
| `esh-server` | 10.0.50.45 + 10.100.50.40 | `…:4411:b105:50:45` |
The v4 secondary moved from the UDM to the **NH3 AdGuard** — reachable over
Site Magic and authoritative for the zone. ⚠ **A secondary only fails over on
SILENCE, not on wrong answers**: NXDOMAIN is a *successful* answer, the client
accepts it and never retries. A secondary that doesn't know your private zone is
a coin flip, not a spare tyre. `esh-cameras` deliberately untouched — routing
camera DNS through AdGuard's filtering risks their cloud features.
## The naming scheme became real
The resolver address is the scheme's first live use, replacing a MAC-derived
SLAAC address that would break on a NIC change. All three `esh-server` Linux
hosts now carry `4411:B105` ("FOR ALL BIOS"):
```
esh-docker-vm …:4411:b105:50:45 esh-pve-nas …:4411:b105:50:55
esh-vm-db …:4411:b105:50:60
```
Applied by an `if-up.d` hook that **derives the prefix at runtime** (self-heals
on re-delegation), backgrounds itself with a retry (SLAAC may not have landed;
a blocking hook would stall bring-up on a headless box), and adds nothing to
existing config. **Not** an `iface … inet6 static` stanza — on Debian that sets
`accept_ra=0` and would strand the host.
⚠ **Proxmox bridges need `accept_ra=2`.** `esh-pve-nas` had link-local only
despite every sysctl looking right: `vmbr0.forwarding=1`, and the kernel ignores
RAs on a forwarding interface unless `accept_ra` is explicitly `2`. Fixed with
`accept_ra_defrtr=0` alongside, so it takes the prefix but **declines the default
route** — an IPv6 identity with no change to a hypervisor's routing. Expect this
on every Proxmox node when its LAN gets v6.
Canonical: `docs/pfi/ipv6-naming-scheme.md`. UniFi limits:
auto-memory `reference_unifi_dns_rdnss_limits`.
@@ -1,308 +0,0 @@
# `[2026-08-24]` Homepage: remote-label consumption re-verified, then the board relaid out on a uniform grid
Prompted by the operator: *"Homepage on esh-vm-docker lists remote dockers and
can absolutely consume their labels, please verify again. I am still
unsatisfied with the layout and aesthetics."*
## The verification — the operator was right, and the record now says so
**Homepage on `esh-docker-vm` discovers services by container label from all
five Docker engines in `conf/docker.yaml`, not just its own.** This is not an
inference; `GET /api/services` returns every card's `server` field, and the
2026-08-24 snapshot resolves to:
| `server` | host | label-discovered services |
|---|---|---|
| `ana-pfi-docker` | 10.250.50.70 | 30 |
| `irv-ml1-docker` | 10.100.79.3 (over WireGuard) | 15 |
| `ana-ml2-docker` | 10.250.50.54 | 14 |
| `esh-vm-docker` | 10.0.50.45 (the dashboard's own host) | 13 |
| `nh3-pfi-docker` | 10.100.50.40 | 2 |
**74 of 107 cards are label-discovered, and only 13 of those come from the
dashboard's own engine** — the other 61 are read off four remote hosts,
including irv-ml1 across the WireGuard tunnel. The remaining 33 carry
`server: null`: those are the manual `services.yaml` entries — hardware, BMCs,
hypervisors, printers, and user-level systemd services that have no container
to label in the first place. **That null is the only thing "not label-driven"
about this dashboard**, and it is a property of the entry, not of the host it
points at.
⚠ If a future session doubts this again, the check is one command and takes two
seconds — do not reason about it from the docs:
```bash
curl -s http://10.0.50.45:5100/api/services \
| jq -r '.[].services[] | .server' | sort | uniq -c
```
## What was actually wrong with the layout
Measured with Playwright against the live board (per-group `card=` width, card
height spread, and a geometric title-vs-status overlap test), not judged by
eye:
- **Card width changed at every group boundary.** `columns:` is not a density
dial — it sets `lg:grid-cols-N` for one group, so it fixes that group's card
width. Notes rendered a single **1464px** card; News and Media **728px**;
Eval & Retrieval **286px**; everything else 360px. Scrolling the page, the
grid resized five times.
- **Long names printed underneath their own status pill.** Measured by
re-injecting the old rule and testing the title text node's box against the
status cluster's box: **6 cards, all on the AI tab** — 3 in Inference, 2 in
Dormant, 1 in Eval & Retrieval; zero on the other three tabs, which is why
it survived earlier passes. Root cause is a genuinely counter-intuitive one:
the rule reserved a
78px gutter with `padding-right` and relied on `overflow: hidden` to hold it,
but **overflow clips at the padding box, not the content box** — so the
reserved gutter was spill room the title rendered straight through. The
intended `text-overflow: ellipsis` never fired either, because the ellipsis
is painted by whichever block's own line overflows, and here that is the
anonymous box around the bare title text node, which does not carry
`overflow`.
- **`AI Systems` / Scriberr was on all four tabs** — the 2026-08-18 UltraSeedbox
bug recurring, this time arriving from a container label rather than from
`settings.yaml`.
- **Icons were grey smudges.** Homepage masks every glyph over
`--color-logo-start/stop`; stock slate-400 → slate-700 sinks the bottom half
of each icon into the card fill.
- Bookmark groups and Jellyfin's trailing stream rows were the two components
the theme had never reached.
## The fixes
`stacks/homepage/conf/settings.yaml` — **all 20 groups to `columns: 4`.**
`stacks/homepage/theme/australis.css.in` → rebuilt → `conf/custom.css`:
gutter held by wrapping, description clamped to 3 lines (floor still 2), icon
ramp overridden, bookmark + trailing-widget components themed, group gap
10px → 22px. `stacks/scriberr/compose.yaml` — `homepage.group` → `AI - Audio
Tools`, container recreated on ana-ml2.
After: **every group renders at card=360**, and the collision count is zero.
Before/after, all four tabs: `http://10.100.10.50:8090/b/homepage-relayout/`
(24h TTL; also on the standing link board).
## ⚠ Three traps worth carrying forward
1. **"Columns = member count" is RETIRED** (it was the 2026-08-18 rule). It was
avoiding dead cells in a short last row and bought a worse defect. A short
last row is what a grid looks like; a card wider than its neighbours is what
a mistake looks like.
2. **A `:root` override of a Homepage theme variable is silently ignored.**
Homepage sets `--color-logo-*` on `.theme-slate`, and that class is on the
`<html>` element — the same element `:root` matches. `.theme-slate` (0,1,0)
beats `:root` (0,0,1), so the override does nothing and looks like the
variable is not the one in play. `html[class]` (0,1,1) wins, and does not
hard-code which `theme-*` class is active. Specificity alone is not enough
either: a custom property resolves from the *nearest* ancestor that sets it,
so the override has to land on `<html>`, not on `<body>`.
3. **The post-recreate tab-bar loss is INTERMITTENT, not guaranteed.** The
2026-08-19 note reads as though every recreate costs up to an hour of broken
render. This recreate came up correct within 10 seconds — fresh payload on
the first poll, all four tabs clickable a minute later. Recreate, *check*,
and only then walk away if it is actually in the broken state.
Also re-confirmed, since the change depended on it: **a `settings.yaml` edit
needs a container recreate, not a restart.** `docker restart homepage` left the
old `"columns":1` payload embedded in the served HTML with the correct file
mounted and readable inside the container; `compose up -d --force-recreate`
cleared it immediately.
## Deliberately not done — operator's call
The Main tab still opens on three sparse bands: **Notes** (1 member) and
**Games** (1) each burn a full 4-wide row, and **News** has 2. Merging Notes +
News, or folding Games into Apps, would tighten the top of the page — but that
is information architecture, not layout, and the group names are the operator's.
Surfaced rather than done.
→ **Resolved in pass 2 below**, where the operator delegated the naming
("re-categorize however you want"). Notes + News became `Daily`, Games folded
into `Apps`, and the `AI - Audio Tools` placement in this pass was superseded
(Scriberr moved on to `AI - Studios`).
---
# `[2026-08-24, pass 2]` Recategorised on "do I open this?", API groups collapsed
Operator, after seeing pass 1: *"You can re-categorize however you want.
service networking tab is uneven, you can split out the adguard cards, etc.
most of the issues are that tools I use and have a UI are interspersed with API
endpoints which are largely informational only. They might even go in their own
cards or start collapsed."*
## The axis
Every group is now either **tools** (expanded, top of tab) or **endpoints** (an
API, a broker, an agent — `initiallyCollapsed: true`, bottom of tab). A
collapsed group still renders its eyebrow and rule, so presence costs one line
instead of two rows.
Second, quieter rule that fell out of the same pass: **a group's members should
all carry a widget or none should.** A stat strip adds ~50px, so one widget card
in a row of plain ones opens a void under the plain ones — which is most of what
made the 13-member `Service Networking` band look broken.
## Shape
- **Main** — `Daily` (Memos, Miniflux, Nevermore, SearXNG — replaces the
1-card Notes and 2-card News bands), `Monitoring`, `Apps` (12; absorbed the
1-card `Games` band), `Media`, `UltraSeedbox`.
- **AI** — `AI - Gateways & Chat` (8) and `AI - Studios` (6) expanded; then
`AI - Inference` (7), `AI - Eval & Retrieval` (4), `AI - Speech (TTS)` (4),
`AI - Audio Tools` (2), `AI - Dormant` (6) all collapsed.
- **Toolchain** — `DNS & Filtering` (3), `Reverse Proxies` (2),
`Compose Consoles` (5), `Toolchain` (3), `Agents (no UI)` (6, collapsed).
- **Infrastructure** — unchanged; every card there is already a console.
Measured after: every group `card=360`, and `DNS & Filtering` and
`Reverse Proxies` both `h=134..134` — dead flush.
## ⚠ The move that made it affordable
**The sixteen GPU-backed model seats were NOT relabelled.** `homepage.group` is
read at container **creation**, so renaming `AI - Inference` to something
clearer would have meant recreating six vLLM seats plus four eval seats plus
four TTS engines — multi-minute model reloads on endpoints peers reach through
the gateway. Order plus `initiallyCollapsed` buys the same separation for free,
so the names stay ugly on purpose. **Do not spend that recreate on a label.**
28 containers *were* relabelled — all cheap web services — via five rerunnable
elway playbooks, `playbooks/homepage-regroup-<host>.yaml`. The canonical
`stacks/` tree was synced to match afterwards, so intent and reality agree.
`initiallyCollapsed: true` is a per-group key in `layout:`; confirmed present in
this build (`defaultOpen: !(group?.initiallyCollapsed ?? global)` in
`/app/.next/server/pages/index.js`).
## AdGuard (ANA) gained its widget, and the credential is fleet-wide
It was the only AdGuard without a query/blocked/latency strip, so it sat short
beside two tall siblings. **One `infra-ops` AdGuard login authenticates against
all three instances** (ANA `:8053`, NH3 `:8080`, ESH `:8080` — all returned 200
on `POST /control/login`, verified 2026-08-24). Vaulted at
`secret get nh3-dev/adguard-infra-ops-password`; written to
`/opt/docker/compose/adguard-ana/.env` (0600, root) and never into git. Its icon
was also the odd one out (`mdi-dns` against two `si-adguard`).
## ⚠⚠ `initialSettings":{}` — the tab-bar mystery is a SWALLOWED EXCEPTION
The biggest durable finding of the day, and it cost ~25 minutes. Full write-up
in `stacks/homepage/README.md`; the short version:
`initialSettings":{}` in the served HTML is **the catch branch** of the page's
data loader, not a warm-up and not a cache. And the error can vanish without
trace: the logger is assigned as the first statement *inside* the same `try`,
and the `catch` only logs `if (logger)`. If the logger is what threw, nothing is
written anywhere — which is exactly what was observed.
Ruled out by measurement, do not re-run: `/api/services`, `/api/bookmarks`,
`/api/widgets` and `/api/hash` all return **200 with correct content** while the
page serves `{}`; restoring the previous known-good `settings.yaml` reproduces
it identically; `/api/validate` returns `[]`; disk and permissions are fine.
**One-command test:**
`curl -s http://10.0.50.45:5100/ | grep -o 'initialSettings":[^,]\{0,20\}'`
**What broke the streak:** three consecutive recreates came up empty, then
rolling the 8.6 MB `conf/homepage/logs/homepage.log` aside and recreating healed
it within 15 seconds. That is one observation, not proof — but it is a coherent
mechanism (oversized log → logger init throws → silent catch) and it is the
cheapest thing to try first next time.
---
# `[2026-08-24, pass 3]` Rebuilt on Australis Skyfall — dual theme, light shipped
Operator supplied the Skyfall design-system README and said "Go full with
skyfall."
## The bundle was already in this repo's git history
**The Skyfall tokens did not need to be hunted down.** A predecessor vendored
them on 2026-08-19 and a later commit deleted them; git kept everything:
```bash
git show 45c1995:stacks/homepage/theme/colors.css # 208 lines, BOTH themes
git show 45c1995:stacks/homepage/theme/layout.css # calm-depth tokens
git show 45c1995:stacks/homepage/theme/typography.css
git show 45c1995:stacks/homepage/theme/fonts/Supreme-{400,500,700}.woff2
```
`colors.css` carries `:root` (dark) **and** `[data-theme="light"]` (Skyfall
Day) — so the light ramp is canonical, not derived. That killed the entire
objection from the previous answer, which was correct only about the
`australis-design` skill ("Always dark first. No light mode in this system").
**Skyfall is the dual-theme derivative; australis-design is the terminal
theme. They are different systems and only one of them has a light mode.**
## ⚠⚠ REMOVING `theme:` FROM settings.yaml BREAKS THE DASHBOARD
The documented way to get Homepage's own light/dark toggle is to leave `theme:`
unpinned. **Do not.** With the key absent, the page's data loader throws and its
catch branch serves `initialSettings: {}` — no tab bar, no layout, no i18n.
Measured, not inferred: six force-recreates over seven minutes all came up
empty with the key removed; restoring `theme: dark` rendered correctly on the
next recreate in **12 seconds**. `/api/services` stays 200 and fully correct
throughout, which is exactly why this reads as a caching or warm-up problem and
is not one.
This is the first *confirmed* trigger for the long-running "tab bar goes
missing" mystery. It does not explain every occurrence (the symptom has
appeared with `theme:` present), but it means **the first diagnostic step is
now `git log -p -- stacks/homepage/conf/settings.yaml`**, not container
archaeology. Also retires an earlier lead from this same session: rolling the
8.6 MB `homepage.log` aside once coincided with a recovery, but did nothing
during the `theme:`-key episode — coincidence, not cause.
## So the toggle is ours
`conf/custom.js` renders it (was an empty placeholder). Precedence:
1. explicit choice — `localStorage['skyfall-theme']`, written by the toggle;
2. OS preference — `@media (prefers-color-scheme: light)`;
3. dark — Skyfall's default.
`theme/build.py` re-emits each vendored `[data-theme="light"]` block twice: as
`[data-theme="light"], html.light`, and inside the media query scoped to
`html:not([data-theme="dark"]):not([data-theme="light"])`. **That `:not()` pair
is what lets a stored *dark* choice survive a light-mode OS.** Verified across
both OS preferences: load, click, click again, reload — all four correct.
⚠ Homepage keeps its own `class="dark scheme-dark theme-slate"` on `<html>`
regardless, because `theme:` is pinned. That is fine and was checked
explicitly: with the dark class present AND `data-theme="light"`, every themed
surface resolves to Skyfall Day, because our rules carry `!important` on the
surfaces Tailwind's `dark:` variants would otherwise claim. **`data-theme` is
the control surface; the class is not.**
## The anti-fork guard is now mechanical
`build.py` records the SHA-256 of each vendored file and **fails the build** on
a mismatch, rather than warning. A vendored file is either byte-identical to
the bundle or it is a fork wearing the bundle's name. Overrides go in
`skyfall.css.in`, which is written entirely against the semantic layer
(`--surface-*`, `--text-*`, `--border-*`, `--success/--danger/--warning`) — no
raw family tokens, no colour literals.
The one place a literal is unavoidable: Homepage consumes
`--color-logo-start/stop` as `rgb(var(--x))`, which cannot take an `oklch()`.
Those four values are exact sRGB conversions of real tokens (`--sea-80`,
`--blue-base` for dark; `--sea-40`, `--blue-deep` for light), computed rather
than eyeballed, with the conversion recorded in the file.
## Deviations, all deliberate and all written down
- **The aurora ribbon under the tab bar is gone.** Skyfall sanctions exactly two
accent expressions — the active rail and hero-only glows — and a decorative
gradient across the chrome is neither. The colour moved to a 2px accent bar
plus `--accent-soft` fill on the active tab, which *is* the rail.
- **Widget stat values moved from the display face to mono**, per Skyfall's
"numbers and telemetry are always `--font-mono`".
- **Two font substitutions**: Space Grotesk for Bespoke Sans, JetBrains Mono
for Victor Mono. Only Supreme was ever vendored, and Skyfall's own notes call
Victor Mono "user-supplied". Two-line swap when the real faces arrive.
Dark + light, all four tabs: `http://10.100.10.50:8090/b/homepage-skyfall/`
@@ -1,46 +0,0 @@
# [2026-08-24] Scriberr transcription deployed on ana-ml2, GPU1
Self-hosted audio/video transcription + diarization. Operator chose GPU
placement over ana-docker (8 cores shared with 50 containers, 37 GB disk)
against ana-ml2's 96 cores, `/tank`'s terabytes and GPU1's headroom.
**Live:** `http://scriberr.ana.internal:8080` (DNS alias added), health `healthy`,
all seven backends up, zero failures: `whisperx pyannote sortformer parakeet
canary voxtral openai`. ~30 GB of weights on `/tank`.
Stack: `stacks/scriberr/`. Full gotcha list in auto-memory
`reference_scriberr_ana_ml2`.
## Three upstream bugs, none of them ours
**1. The Blackwell image does not exist.** Upstream's README documents
`scriberr-cuda-blackwell`; GHCR has **no tags for it**. Published
`scriberr-cuda` covers sm_61–sm_89 only — on these sm_120 cards it fails or
silently drops to CPU. The real sm_120 path is `Dockerfile.cuda.12.9`
(CUDA 12.9.1, cu128 torch), **built from source**. Do not "simplify" the compose
back to the published image.
**2. It must run as uid 10001, not 1000** — and the error lies:
`unable to open database file: out of memory (14)`. Error 14 is
`SQLITE_CANTOPEN`, not an OOM, on a box with 566 GB RAM. That Dockerfile creates
`appuser` at 10001 (Ubuntu 24.04 owns uid 1000 as `ubuntu`) and chowns `/app` to
it, while the entrypoint's PUID remap covers only the data dirs.
**Isolated by elimination**: SQLite writes fine to `/tank` as 1000 → not the
mount; fails on a plain named volume too → not the storage; the **published CPU
image works at PUID=1000** because there `appuser` *is* 1000.
Generalisable: *when a container "permission" bug appears, compare the uid the
image was BUILT for against the uid you are RUNNING as.*
**3. `UV_LINK_MODE=copy` is required.** Scriberr builds each backend's Python env
with `uv` at start; uv's reflink mode fails on overlayfs+ZFS with
`Failed to clone … Resource temporarily unavailable (os error 11)`. **Partial
failure** — WhisperX and PyAnnote came up and the app looked fine while Parakeet
and Sortformer were silently absent. Occurrences 2 → 0 after the fix.
## Related
`speaches` on irv-ml1 **stopped** the same day (stack retained, one command to
restart): Eyra was abandoned pre-implementation because Scriberr covers the need,
leaving it with no consumer. Scriberr runs its **own** WhisperX in-container and
is **not** a speaches consumer. Idle footprint at stop was 274 MiB, not the
~5.9 GB quoted — that figure is the loaded-model working set.
@@ -1,92 +0,0 @@
# ERP/RP tune run-01 COMPLETE — 7.36h, gate passed on the axis it was built for
`[2026-08-25]`
## The run
1312/1312 in 7:21:52 train_loss 2.793 epoch 1.0
20.1 s/it FLAT across every 100-step window (round 1: 35-46.5 s/it)
adapter: /tank/erp-tune/run-01/adapter/ 410 tensors, provenance.json
**47% faster than the round-1 projection of 13.85h**, from two changes: the
bucketed sampler and flex attention. Rate was flat — 19.7 / 19.8 / 20.4 / 20.3
across the four 100-step windows — which means the 35-46.5 spread in round 1 was
*entirely padding*, and removing padding removed the variance rather than just
the mean.
⚠ **I quoted three different ETAs (6.9h, 8h, 7.3h) before I started using a
rolling average.** The first two were instantaneous tqdm readings off a number
that swings 17-25 s/it with batch width. Only the rolling rate was honest. Same
measure-don't-sample discipline I wrote into the throughput playbook, violated on
the one metric I kept reporting.
## lora_B gate — PASSED, twice
checkpoint-100 205/205 non-zero, median norm 0.829
final adapter 205/205 non-zero, median norm 1.708
vision_tower tensors: 0 on both
Median norm rising 0.829 -> 1.708 means it kept learning through the whole run
rather than saturating early. This check **never ran in round 1** (died at step
19, first checkpoint was 100) and it is the only failure mode that stays
invisible until the acceptance gate reports base-identical numbers.
## The gate — brokkr-smithy-dev
**It did the thing it was built to do:**
metric base A/B tuned delta floor
attractor hit rate 94.8% / 96.8% 84.5% -11.3pt 2.0pt
diversity (pairwise) 0.213 / 0.221 0.3948 +0.178 0.008
Diversity moved **22x its own noise floor**. Attractor rate (how often the model
reaches for the same names and phrasings) fell 11 points against a 2-point floor.
T1 100 · T2 95 · T3 96-97 · T4 98 · T5 100 · T6 81-82 · core ~94.2
memorisation: 0.0000 on BOTH arms, all three corpora
**Zero memorisation closes the R20 licensed-prose exposure on measurement rather
than argument.**
⚠ **Caveat brokkr volunteered rather than buried:** the tuned arm lost 18 of 192
generations to truncation/degeneracy against base's 1-2. Lopsided exclusions
plausibly flatter the diversity magnitude. Direction is unambiguous at 22x floor;
the number carries an asterisk.
## The noise-floor near-miss — the methodology lesson
brokkr was one step from reporting a 13-point T6 regression **that sat inside
twice his instrument's own variance.**
--per-type 32 max swing across tasks: 9 points
--per-type 128 max swing across tasks: 1 point
His gate criterion is "no task regresses by more than one item" = 3.1 points at
n=32. **The instrument's own run-to-run noise was 3 items.** He was scoring a
preregistered gate at 4x finer resolution than it could resolve, and caught it by
running a control he did not strictly need. Quadrupling n collapsed the noise
exactly as binomial statistics predicts.
⚠ **Root cause of the noise is a property of the SEAT:** `max-num-seqs` is unset,
so with a 218,625-token KV cache the scheduler batches freely up to vLLM's
default of 256. Continuous batching changes reduction order and borderline items
flip. Temperature 0 buys deterministic *sampling*, not deterministic
*arithmetic*. He declined a `--max-num-seqs 1` determinism control for the right
reason: a floor measured on a seat serving one request at a time is not the floor
that applies to the seat we ship.
## The confound I built and he caught
I optimised a pipeline for production and then handed him its output as an eval
instrument **without asking whether those were the same job.** The tuned arm
would have reached the seat as NVFP4A16 while his base arm was bf16 — any
regression would have been tuning-damage OR quantization-damage with no way to
separate them, and the gate's whole question is "did the tune cost us
capability."
**Both arms now bf16, same seat, same port, argv differing in exactly two
lines** (weights path, served name), template sha256 identical
(`ae53464bf3be2580`), KV cache identical to the digit (218,625 tokens across all
three launches). Quantization moved *downstream* of the gate.
See [[2026-08-25-refusal-retention-probe]] for the axis his gate did not have.
@@ -1,86 +0,0 @@
# The 8.6% MFU was an accounting artifact — attention on Ampere kernels
`[2026-08-25]`
## The answer
**Real utilisation was 17-20%, inside the honest stock band.** The 8.6% divided
the *intended* (windowed) FLOPs by the wall time the *dense* reality took.
nominal billed 27.1 TFLOPS x 34.85 s = 9.4e14 FLOP
dense-sliding extra 25 layers, 2 seqs, 4 passes = +8.2e14
padded full layers lose the causal skip = +3.5e14
work performed ~ 1.8e15 = 51-61 TFLOPS
The card was doing ~2x the arithmetic the architecture specifies, and the excess
was the sliding window being computed and thrown away.
## Two independent methods agreed
scaling fit (3 points, 2 params, residuals <3ms over 8x range)
A = 6.87e-4 s/token B = 8.85e-8 s/token^2
quadratic share: 20.9% @ w=2048 -> 67.8% @ w=16384
kernel table (device rows only)
attention 22,835.8 ms 65.2% fmha_cutlass*_sm80
dense GEMM 2,774.0 ms 7.9%
other 5,739.0 ms 16.4%
**67.8% vs 65.2% — 2.6 points apart, no shared assumptions.** The two-term fit
needed no constant term, which refutes launch-bound outright (3,840 expert-GEMM
launches per forward are not the cost).
## The mechanism, source-verified by brokkr's panel (arm: Bil)
masking_utils.py:292-301 _ignore_causal_mask_sdpa requires
kv_length < local_attention_size. 16384 >= 1024,
so THE SLIDING MASK ALWAYS MATERIALISES.
sdp_utils_cpp.h:259-267 flash rejects ANY explicit mask
sdp_utils.cpp:647 cuDNN head_dim capped at 128 -> unreachable
Context.h:480-485 prefer-cuDNN needs major 9 or 10; sm_120 is 12
⚠ **The kernels are `sm80` — Ampere-generation CUTLASS on a Blackwell card**,
with the forward on `gmem`, the memory-efficient backend's slowest fallback tier.
## What actually fixed it
**Bucketing (bucket-to-pair, shuffle-to-mix)** — 29.9% padding -> 0.0%, and
78.3% of micro-batches become exactly zero-pad, which puts the 5 global layers
back on `is_causal`. Measured: padding costs **9.4% MORE time for 24% LESS work**
at fixed width, because an explicit mask knocks those layers off the fast path.
⚠ **Bucket size is NOT a diversity knob.** Swept across a 256x range, roots per
accumulation window stayed flat at 3.54-3.61. The global micro-batch shuffle does
all the mixing; the bucket only costs padding. Use the tightest bucket.
**flex_attention** — Triton-generated so it compiles for sm_120 instead of
shipping sm_80 binaries. 21.7x on sliding layers, 2.1x on global. Needs mandatory
`kernel_options` at 32x32 blocks: 64x32 needs 102,400 bytes against a
**101,376-byte hardware ceiling** — misses by 1 KB, and Triton is already opting
into the full 99 KB, so it is the card, not a default.
## ⚠⚠ The trap that produced TWO wrong published conclusions
`torch._dynamo` defaults to a recompile ceiling of **8**. Every distinct sequence
width is a new shape. On hitting the ceiling dynamo does not error — it silently
falls back to UNCOMPILED flex, which is ~20x slower AND documented to *"not work
with the backwards pass and may produce incorrect results."*
That artifact produced a bogus **0.76x slowdown** and a bogus **2.9% loss
divergence**, and I believed and reported both. Raising the limit to 256 flipped
the speed result to 1.41x.
The loss divergence turned out to be real but benign — adjudicated against fp32
MATH ground truth, both backends sit ~2e-3 from truth with flex fractionally
CLOSER at every width. **Do not re-open it by comparing the two backends to each
other; that cannot answer it. Compare to fp32.**
## Process lesson
brokkr's panel produced **four self-retractions in ninety minutes**. Every
retraction was a derivation; every survivor was a measurement. And the whole
head_dim-512 SDP problem was **already documented in zerofata's published Axolotl
config since April** — the right first stop for "why is this architecture slow"
is practitioner configs for that exact base, before any panel.
Playbook: `docs/pfi/training-throughput-playbook.md`, commit `7b5fd91`.
@@ -1,84 +0,0 @@
# NVFP4A16 serving pipeline — built, validated, and the MoE landmine it found
`[2026-08-25]`
Pipeline at `scripts/erp-tune-serve/` (commits `6a85829`, `ab980e9`).
Validated end-to-end against checkpoint-100 before the real adapter existed.
## ⚠⚠ THE LANDMINE: a `targets=["Linear"]` recipe misses EVERY MoE expert
before linearize_moe: 427 Linears, 205 targeted, experts 0
after linearize_moe: 11,947 Linears, 11,725 targeted, experts 11,520
(30 layers x 128 experts x 3 projections)
Gemma-4 stores each layer's 128 experts as two fused 3-D `nn.Parameter` tensors
(`gate_up_proj` [128,1408,2816], `down_proj` [128,2816,704]) — note the absent
`.weight` suffix. A Linear-targeting recipe resolves 205 of 427 modules and
**zero experts**, leaving 22.84 B params (88.5% of the model) in BF16 with no
warning.
**This is the same defect that killed QLoRA here via bitsandbytes.** The blind
spot is in the *checkpoint layout*, not the tool. Fix:
`llmcompressor.modeling.moe.linearize.linearize_moe` — no registration needed,
Gemma-4 satisfies `FusedExpertsProtocol` structurally. Playbook §3.15.
## Scheme: NVFP4A16, deviating from the playbook default, on measured grounds
brokkr benched the W4A4 quant of this checkpoint at **12% on contradiction
detection with CoT off against gen's 81%** — the signature of 4-bit input
activations on a reasoning-dense task. Plus W4A4 KLD is 2-4x worse past ~10k ctx
on sm_120. This is a 16,384-ctx RP seat. Marlin's prefill cost accepted.
⚠ Several HF repos named `…-NVFP4A16` declare `input_activations num_bits 4` —
W4A4 wearing an A16 label. The script refuses if the emitted config says 4.
## Four silent defects the dry run found
1. **transformers 5.15 MIGRATES the config schema on save** — drops
`global_head_dim`/`num_global_key_value_heads`, writes `per_layer_config`.
transformers 5.10 (the llmcompressor venv) then reads `num_key_value_heads`
as None and dies with `TypeError: unsupported operand type(s) for //`.
Every working artifact on the box uses the OLD schema. Merge now downgrades it.
2. **llmcompressor cannot auto-init a processor for a multimodal checkpoint** —
pass the tokenizer explicitly as `processor`.
3. **`save_pretrained` does not carry `processor_config.json`** — vLLM then fails
with "Can't load feature extractor", which reads as a vision bug.
4. **The quant needs more than GPU1's free 32 GiB.** `quant_with_gen_down.sh`
stops `vllm-gen` and restores it from a trap on EVERY exit path, using
`docker start` not `compose up` so the container returns with its exact config.
## Verified on the emitted artifact
49 GB -> 17 GB, format nvfp4-pack-quantized, a=null (genuine A16)
weight_packed 11,725 of which expert 11,520
tokenizer truncation: clean (§3.14 trap avoided by calibrating on the
encode cache, so the tokenizer is never called
with truncation=True at all)
served: Marlin NVFP4 kernel + Marlin MoE backend, coherent generation
⚠ The reference `nvfp4a16` artifact triggers a vLLM warning that q/k/v carry
*different* weight global scales ("likely reduced accuracy"). **Ours does not** —
llmcompressor 0.12 links weight observers across fused groups automatically. The
in-house quant is better than the downloaded one on that axis.
## ⚠ MERGED WEIGHTS ARE MANDATORY — and not for the reason we assumed
The open question was whether LoRA-on-NVFP4 hot-swap still silently no-ops.
Retested on `vllm/vllm-openai:latest`: **it refuses to start.**
AttributeError: To support LoRA for MoE model,
'get_expert_mapping' must be implemented
The check is in `vllm/lora/utils.py::process_packed_modules_mapping` and branches
on `is_moe_model()` — **quantization is not in the condition.** `gemma4.py`,
`gemma4_mm.py`, `gemma4_mtp.py`, `gemma4_unified.py` all have ZERO occurrences;
`deepseek_v2`, `mixtral`, `glm4_moe`, `ernie45_moe` implement it.
**vLLM cannot serve a LoRA on ANY Gemma-4, bf16 or quantized.** Merging is the
only path for this architecture, and it would have bitten identically on the
unquantized base. A loud refusal is strictly better than the 0.24.0 silent no-op,
which shipped a base model wearing the tune's name.
⚠ Base-viability pre-flight is now playbook §3.11 — three greps before picking a
base. **Grep the CLASS, not the file**: `mistral.py` greps as `SupportsLoRA=0`
and is fully LoRA-capable via inheritance from `LlamaForCausalLM`.
@@ -1,64 +0,0 @@
# Refusal retention — the axis the gate did not have, and the axis I measured wrong
`[2026-08-25]`
## Why it exists
brokkr's gate measures reasoning (T1-T6), craft (diversity/attractor) and
regurgitation (memorisation). **Nothing measured whether the model still
COMPLIES** — which for this seat is arguably the most important property.
The risk is specific to our operation order. We do **tune(abliterate(stock))**,
so the tune has 57.7M tokens of opportunity to walk the abliteration back. *A
tune that gains 41 items of contradiction detection and quietly re-installs
refusals is a failed seat that passes the entire gate.*
## The measurement — controlled, single instrument, both arms
arm HARD DEFLECT COMPLY
base 0/100 0 100
tuned 29/100 0 71
Same seat, same probe, temp 0, `mlabonne/harmful_behaviors` x100.
Probe: `scripts/training-probes/refusal_probe.py`.
**The tune added 29 general-harm refusals where the base had none.**
Two things fell out:
- **The instrument validates.** Base measured 0/100 on my generated-text regex
against Heretic's recorded 3/100 from a first-token-probability scorer. 0 vs 3
is agreement — the incomparability worry was right caution about a non-problem.
- **DEFLECT is 0 on BOTH arms, so the free control fires.** An instrument
artifact does not care which arm it runs against. Both zero means the model is
**binary** — refuses in refusal-language or engages, no soft-deflection tail.
The R19 undercount does not apply here.
## ⚠⚠ But it is the WRONG AXIS — brokkr's catch, and it is the better one
`mlabonne/harmful_behaviors` is **general harm** — weapons, malware, fraud. **The
abliteration was not run so the model would explain bomb-making. It was run so
the model would engage with explicit fiction.** Different refusal surfaces; a
model moves on them independently.
I picked that set because it was cached, had a recorded baseline, and was what
the abliteration tool used. **Every one of those is a reason it was convenient,
not a reason it was right** — and "it has a baseline" was actively misleading,
because a comparable number for a question nobody is asking looks like evidence.
**29/100 general-harm refusals on a seat writing prose the operator was actively
praising is plausibly the DESIRED shape**, not a defect. General-harm refusals
returning while domain compliance holds is close to ideal for an internal
creative seat. I would have reported it as damage.
**The load-bearing cell is COMPLY 71, not the 29.** Stock refused 100/100;
anything near that would mean the abliteration was undone. 71 complying means
"partially walked back on one axis" — a different finding, and only one of the
two threatens the seat.
Domain-compliance probe (the right axis, from R19's track-2 map) is brokkr's,
pending. Scaffold supplied: `scripts/training-probes/counted_classifier.py`
(`2a05ae9`) — classify-never-surface, three-way, ERROR path deliberately does not
log the exception body because an exception can echo the prompt back.
Playbook §3.13. See [[2026-08-25-erp-tune-run2-complete]].
@@ -1,123 +0,0 @@
# Worldtree b188 + b189 bridge cutover, and the selene metadata that lied
`[2026-08-25]`
Three arcs in one day, all infra-ops side, all landed.
## b188 — matrix.yaml pre-sync (#406/#409/#410 closed)
From b188 the bridge reads per-agent `rendering` + `ambient_buffer_size` from
`config/matrix.yaml` ONLY; agent `config.yaml` matrix blocks are gone from the
image. Staged as `6417115` in `worldtree-instance-configs`, deployed to both
instances with operator approval.
- mimir gets thinking-to-thread + tool-call reactions + 7 reaction labels;
forseti and lofn stay quiet.
- **Rider #409 pruned six dead agents** (bragi, leif, troi, soong, cara, glados)
from BOTH rosters — originally scoped personal-only, which I flagged as a
possible oversight and it was. **The settling fact worth keeping: the engine
roster comes from the image's baked `agents/` directory (only `config/` is
bind-mounted), and both instances run the same image**, so instance-level
evidence about which agents the engine lists generalises by construction.
- Both rosters now exactly `[mimir, forseti, lofn]` — the three engine agents
actually bridged. mask/vili/echo exist in the engine, deliberately unbridged
(operator ruling).
⚠ **Edited text-surgically, not via a yaml round-trip** — PyYAML would reflow
1,249 lines and drop every comment, and the comments are the documentation.
⚠ **`deploy-wt-config` uses `docker restart`, NOT `compose up`.** A `compose up`
on corviduo-dev re-resolves the image tag and can silently swap the running
build — which would turn an "inert pre-sync" into an unintended image roll on two
live instances. That property is easy to lose in a future refactor of the script.
## b189 — #407 bridge extracted to its own repo (#404 umbrella closed)
Bridge now `gitea.phasefinal.com/pfi/wt-matrix-bridge`, its own repo, own CI.
⚠ **It publishes to the `pfi` ORG, not `vh`, and the reason is structural:** `vh`
is a **USER**, not an org. Gitea scopes user-namespace packages to the owning
user — there are no package collaborators on a user namespace. **No service
account can ever publish to `gitea.phasefinal.com/vh/*`.** claude-bot is an Owner
of `pfi`, so that is where it goes. Token `wt-matrix-bridge-ci` (id 28, scopes
`write:package,read:repository`), vaulted at
`nh3-dev/.config/claude-bot/gitea-token-wt-matrix-bridge-ci`.
**Minted a dedicated token rather than reuse `claude-bot-sdk-ops` or `arbo-ci`,
both of which already carry `write:package`** — a shared credential cannot be
revoked without collateral. ⚠ A first mint attempt succeeded then failed to save;
Gitea returns a token value exactly once, so it was unrecoverable. Deleted the
orphan (id 27) rather than leave a live package-write credential on the account.
**Both instances PINNED** to `f3f8ec902267` (`e90f436`), closing the #410 shape:
`WORLDTREE_IMAGE` was pinned and the bridge was the one drifting service.
⚠ The pin moved from `b178285b1cb5` because a cross-frontier bug-hunt found the
M_EXCLUSIVE **fallback could itself litter unboundedly** in exactly the state it
was written for. **The safety net had the same failure mode as the thing it was
catching** — and my staged window leaned on that fallback.
## #411 — the debug-room failure, diagnosed twice and wrong both times first
My theory: the alias was held by orphaned rooms. **Refuted by a directory probe
returning 404.** The real cause: Synapse's **M_EXCLUSIVE** — an appservice may
only create aliases inside a namespace it has RESERVED, and
`aipa_appservice.yaml` had `namespaces.aliases: []`.
⚠ **I inferred a cause from a symptom that was the RESPONSE to the cause** — the
log's "re-resolving alias" line is the recovery path firing, not evidence the
alias exists. One directory probe settled it and I reasoned instead.
Fixed with operator clearance: added
`regex: '#aipa-debug-[a-z0-9_-]*:matrix\.phasefinal\.com'` (exclusive) at
`/opt/docker/conf/synapse/aipa_appservice.yaml` on **ana-docker** (NOT
`/opt/docker/data/`, which worldtree-dev's issue cited). **My regex was tighter
than the proposed `#aipa-debug-.*`**, which fullmatches only because `.` also
matches the `:` separator and would equally claim other homeservers.
Pre-apply sweep (an `exclusive: true` claim can make Synapse refuse to START):
aliases matching '%aipa-debug%' 0
total room_aliases on the homeserver 1 <- why it went unnoticed this long
rooms created by @aipa-debug 17 <- the litter, confirmed unaliased
Synapse healthy in 40s, both bridges rode through. **The aliased create then
worked FIRST TRY on the next personal recreate** — designed path, fallback never
fired, both reuse mechanisms live (canonical-alias rediscovery AND the room-id
cache).
⚠ The #411 writer census **inverted its own premise**: the api (uid 1000) cannot
write `/app/sessions`, the bridge (root) can — the reverse of the issue text. But
worldtree-dev's reconciliation is better than "backwards": pre-#407 the bridge
ran from the ENGINE image as uid 1000, and the cutover changed the answer
underneath the issue. Both readings were true at their timestamps. **No live
writer exists for that path anyway** (`sessions.path` is a legacy default, #330
moved the tools off it), so the chown is optional future-proofing.
## selene-1-mini-8b — a config that lied about what answers
forseti's fleet sweep found the only genuine residual in
`worldtree-instance-configs`, and it was **live on both boxes, not just drifted
in git.** Fixed in `a77639d`.
Routing was never broken (`model: "chat-judge"` stays), but `display_name` said
"Selene 1 Mini 8B" and the description said "Atla Selene 1 Mini 8B — reward model
derived from Llama 3.1 8B" while chat-judge has been backed by
**qwen3.8-27b-uncensored** since 2026-08-23.
⚠ **It misstated the KIND of model, not just the identity.** A reward model and a
generative judge are different instruments; a consumer reading
"pairwise/likert/binary/scalar reward model" would expect scalar-reward semantics
qwen3.8-27b does not provide.
Verified: `selene-1-mini-8b` -> HTTP 400 (by design), `chat-judge` -> HTTP 200.
**NOT changed:** the catalog key and the `selene-judgment` role — `model_roles.yaml`
binds to that key, so renaming is worldtree-dev's schema call. Flagged that a role
named after a retired model defeats the purpose of role aliases.
⚠ forseti's sharpest finding is for the operator: **`~/.claude/CLAUDE.md` line 502
lists the retired name in the global tools roster.** A broadcast reaches sessions
that already exist; the roster line keeps minting new ones. His file, awaiting his
word. Also: **there is no fan-out primitive on the bus** — a fleet relay is 73
individual posts, recorded as a real gap rather than papered over.
@@ -0,0 +1,8 @@
# ERP run 7 launched on pfi-gx10 (2026-09-08 23:06 PT)
Operator's direct grant `operator-2026-09-08-rnd-run7` (in-session `/goal: launch training run 7 on gx10`, 23:03 PT), pid 599489, 542 steps, ~80 s/it, ETA ~noon 09-09. Runbook `docs/runbooks/gx10-run-07.md`; canonical `scripts/erp-tune-gx10/{run-07-gx10.json,launch-run-07.sh,build_r7.py}`; commits `a34a72f`, `d8f625e`.
- Variable (brokkr `recipe-erp-seat-sft-r7.json`, sha `2972559e…`): an opening-split slot (293 rows, `opening-split/clean-v1`, root `3ab41c08…`) + companion mask (224 entries) UNION lossmask-r3 → `lossmask-r7` (6,106, 1 dup). `survivors-r7` = survivors-r5 + 293 = 8,505. Root/mask/holdout relayed NAS→nh3-dev→box (no NFS on the GX10), sha-verified.
- Free check: held roots identical to run 6 (173 chunked / 30 truncated / 59 lost turns); slot 293/293 fit_whole; two cwm conversations (cwm-2602, cwm-1826) dropped as unfittable because the union mask left them no loss turn — brokkr: "loss moved, not doubled, satisfied by construction". Padding 17.1% (run 6: 0%).
- ⚠ First launch died in preflight: the recipe JSON never landed because a zsh quoting error silently aborted the staging chain. Lesson: stage files with `scp` in their own command; never chain a heredoc-bearing python inside a single-quoted ssh string.
- Gate after the adapter: brokkr commits `tools/base_window_r7.sh`, `tools/tuned_window_r7.sh`, `tools/verdict_run07.py` before it lands; serve `erp-seat-base-ara` (floors re-locked) → `erp-tune-v7`; cue-length probe as a conformance leg on both arms; DONE sentinel gated on every step's artifact.
@@ -0,0 +1,11 @@
# erp-tune-v6-nvfp4a16 on ana-ml2 :8021, aliased to `trial`, and the Gemma-4 tool-calling findings (2026-09-08)
**Seat.** Operator: "quant the latest trained model into nvfp4 and serve it on ana-ml2 while we train a new model on the gx10." Stack `stacks/erp-seat` (recipe = gemma4-charrp's; true served name only; GPU1, util 0.35, 32K ctx). Artifact `/tank/aimodels/erp-tune-v6-nvfp4a16` (16 GB, compressed-tensors nvfp4-pack, W4A16, 252 ignores incl. 60 router + 191 vision) from `/tank/aimodels/erp-tune-v6-bf16` (merged-run06 relayed gx10→nh3-dev→ana-ml2 in 17 min at ~47 MB/s — no key path gx10↔ana-ml2). Quant pipeline `services/erp-seat-quant/` replicates `prithivMLmods/gemma-4-26B-A4B-it-NVFP4A16`: `linearize_moe` first, assert 11,520 expert Linears, routers/vision/audio ignored, post-steps restore processor configs + template and reset the tokenizer cap. DATA-FREE (~90 s) — playbook §3.16. Commits `911ff20`, `8512dd4`.
**Alias.** Operator: "alias erp-tune-v6-nvfp4 to trial, please. no gate." The `trial` model lives in `stacks/litellm/conf/config.yaml` → `/model/update` refuses it ("Model in config"); deploy `--conf` + `sudo docker compose restart litellm` on ana-docker. Verified 3/3. Commit `48aa2c6`. Block states the seat as unrated on every safety axis.
**Tool calling (operator: "fix toolcalling with the trial seat").**
1. `tool_choice:"none"` → empty turn (content AND tool_calls null, 3/3): vLLM kept the tools in the prompt, the model called one, parsing was off. Fix `--exclude-tools-when-tool-choice-none` (`7f6be8a`). `stacks/gemma4-charrp` has the same exposure, no flag — bouncing it is consumer-visible, operator's call.
2. Forced (named/`required`) tool_choice is PROMPT-DRIVEN on every Gemma-4 seat **by vLLM design**: `vllm/tool_parsers/gemma4_engine_tool_parser.py` sets `supports_required_and_named = False` and `adjust_request` skips the structured-output JSON so the model can emit native `<|tool_call>call:…`. A tune that weakened that syntax honours forced calls only sometimes. Measured (3 conv × 3 turns, real system prompt): v0.26.0 1/9; nightly `311b3513` (v0.27.2rc1, the `gen` seat's image) 6/9 with the round trip clean 3/3 → seat pinned to the nightly (`a34a72f`). tts-dev: 0/18 on v0.26.0 with `gen` 18/18 as positive control; **`response_format: json_schema` 18/18 = the deterministic path**. A parser plugin re-enabling guided JSON would also need JSON extraction in the engine-parser path — not attempted.
3. ⚠ Removing `--reasoning-parser gemma4` does NOT fix forced choice (3/9) and leaks `<|channel>thought` into content 3/3 — keep it and the `enable_thinking:false` pin.
Single-turn probes are FALSE POSITIVES for forced tool_choice — gate on multi-turn with a real system prompt (tts-dev's lesson, confirmed).
@@ -0,0 +1,5 @@
# ESH static WAN follow-ups landed; YTVC chase; gitea webhook repointed (2026-09-08)
**ESH static WAN** `128.177.138.182/30` (gw .181, Cityside Fiber) confirmed LIVE (UDM WAN1 static, uplink up since ~09-05; egress verified; CGNAT at ESH is history). Added to the crowdsec `esh` allowlist on ana-docker. Operator: "land all 3" → (a) FortiGate infra-ops `trusthost3` 23.164.40.160 → `128.177.138.182/32`, verified by a real login from esh-docker-vm (flash backup ran but `execute revision list` errors on the 80F — unconfirmed); (b) dormant `esh-ana` IPsec rebound wan2/192.168.200.111 → wan1/`128.177.138.182` (still disabled); (c) ESH UDM forward UDP 41641 → esh-scale 10.0.50.65 → mesh peers DIRECT via `128.177.138.182:41641` (was DERP lax). Commits `7d5de53`, `a19b7f6`. ⚠ The permission classifier blocked every edge-gateway write until the operator disabled it — three shapes tried, then stopped and surfaced; no tool-laundering.
**YTVC** (operator: "chase youtube voice clipper being down"): (1) every job died at yt-dlp because the nh3-dev danted proxy was retired 09-06 → new fleet SOCKS5 on **nh3-scale** `socks5h://100.64.0.1:1080` (`scripts/setup-nh3-scale-socks-egress.sh`, dante, fleet ACL, residential egress 70.230.226.88; ⚠ dante one-line `{ }` blocks are a syntax error); (2) Homepage href on the dead wg0 IP → `irv-ml1.nh3.internal:8000` (labels live in the UNTRACKED override); (3) full downloads still 403'd = stale yt-dlp 2026.07.04 from a cached Docker layer (`--test` passing while the full fetch 403s = client, not IP/cookies) → `compose build --no-cache api`; (4) the rebuild pulled a whisperx needing `punkt_tab` → staged on the data volume + `NLTK_DATA`. Then yt-voice-clipper-dev shipped v0.3.4→v0.3.6 (cache-bust ARG, punkt baked, **CrisperWhisper 2.0 replaces whisperx**: x7kWJojf1MI → 12 clips with `[UM]`/`[UH]` tags). ⚠ The gitea push webhook had targeted the dead wg0 IP since 09-06 (never fired) → repointed to `10.6.110.50:9008` with the HMAC secret RE-APPLIED (a PATCH without `secret` drops it); test push auto-deployed. Commits `14dd9ab`, `d2df1eb`, `f2b7384`.
@@ -0,0 +1,8 @@
# Run-6 gate: CSAM soft trip, operator adjudication, TRANSFERRED (2026-09-08)
- 22:18 PT brokkr HALTED the run-6 gate: R19's CSAM-drift detector fired at level=review on ONE generation in the tuned refusal k=5 step — metadata only (minor_soft=1 `soft_baby` near anatomy terms, minor_hard=0, sexual=7, salted hash `01724c39…`); the spec (§8) never stores the text and the seat does not log completions, so **the text exists nowhere**.
- Conflict surfaced, not resolved by me: the operator had aliased `trial` → the NVFP4 build ("no gate") two minutes before the halt reached me; brokkr asked that it not be served until adjudication. Left the alias in place (explicit operator ruling), escalated to the operator + Miranda with the one-line rollback offered.
- 22:45 PT operator adjudicated GO, verbatim: "I don't want it regenerated, but I will say that in the vernacular, baby is about the most common pet name you can get, ESPECIALLY during sex. I'm going to adjudicate it as a go. There are unmistakable CSAM terms, but baby is not one of them." Relayed verbatim to brokkr and Miranda.
- brokkr finalized TRANSFERRED without the tuned refusal leg (the cell rule never read it; tool forbids a re-run); tuned refusal telemetry recorded ABSENT for run 6; RESULT carries the adjudication with infra-ops as relay. Operator also CUT the k=25 refusal legs on both arms ("the battery is information enough").
- Cue-length probe (both arms, 288 gens each): tuned 5-word cue → 54/62 words median, 49-word → 114, 221-word → 292; base 168/191 → 209 → 282. Tune's deficit ~120 words at the short cue, ~0 at the long → run-7's lever.
- The choreography honoured hands-off throughout; the base seat served 16:24–21:38 and again 22:45–23:01 PT for the reference arm.
@@ -0,0 +1,15 @@
# ana-ml2 pool health — three actions deferred to a clean-context session (2026-09-09)
Operator ruling 2026-09-09 ~00:30 PT: "snapshot and we'll do all 3 on clean context." Findings commit `3e18a04`.
**Findings (measured 2026-09-09 00:00 PT):**
- `tank` (raidz2, 8× NVMe): ONLINE, **2 CKSUM errors on `nvme7n1`**, boot-time resilver of 638 GB on 2026-09-05 14:26 (box rebooted at 14:26; nvme7 came up late/dirty). No data errors, 58% full. **No scrub since 2026-04-12** — the Debian `zfsutils-linux` second-Sunday cron scrubbed `zroot` on 08-09 but not `tank`; cause unknown (`zpool history tank` shows trims monthly, last scrub 04-12).
- No `nvme-cli` or `smartctl` on the box → nvme7's media-error counter unread.
- `zroot` at **91%** (345 G of 379 G): `docker system df` = images 429 GB (204 GB reclaimable), build cache 74 GB (36 GB reclaimable).
- pfi-pve `NASPool` 7% / `ospool` 19%, scrubbed 09-05 / 08-09, clean.
**The three actions, in order:**
1. `sudo zpool scrub tank` on ana-ml2 (1–2 h of extra I/O; seats keep serving) → on a clean pass `sudo zpool clear tank`; if the scrub finds errors on nvme7n1 → replace path.
2. `sudo apt install nvme-cli` → `sudo nvme smart-log /dev/nvme7` (media_errors, critical_warning, percentage_used) and `nvme id-ctrl` for model/serial; record in the drive inventory.
3. `docker image prune -a`? NO — `docker image prune` (dangling only) + `docker builder prune` on ana-ml2; the 47 unused-but-tagged images need a look first (some are rollback seats: e.g. `vllm/vllm-openai:v0.26.0`, nightlies). Target: zroot back under ~75%.
Also worth a look while there: why the scrub cron skips `tank` (`/usr/lib/zfs-linux/scrub` logic — it skips pools with an active trim/resilver or those not "healthy"?).