Refuse a conf_upgrade that would keep nothing

Merging the raw host.conf.template against a live host conf matched HOSTN_
against HOST1_, classified all 140 real keys as deprecated and would have
removed every credential on the host. Also points the Ollama model default
at a tag that still exists.

Deployment/conf_upgrade.sh

@@ -69,6 +69,22 @@
 # OPERATIONAL SAFEGUARDS
 # ==============================================================================================
 #
+# Unsubstituted HOSTN Placeholders Are Refused
+#   host.conf.template ships HOSTN_ placeholders; a live host conf uses HOST1_ / HOST2_. Key
+#   matching is literal, so merging the raw template against a real host conf classifies every
+#   existing key as deprecated — KEPT 0 — and the install drops every credential in the file.
+#   Verified on HOST1 2026-08-02: it would have removed HOST1_RADARR_API_KEY,
+#   HOST1_EMBY_API_KEY, HOST1_NPM_PASS and 137 others. The template is checked for HOSTN and
+#   the target for a HOST<n> slot; if both are present the run aborts and prints the exact sed
+#   command to substitute. Refused rather than auto-substituted — this rewrites the file holding
+#   every secret on the host, and inferring the slot is not worth being wrong about.
+#
+# Total Mismatch Is Refused
+#   Keeping nothing from a populated conf is never a real upgrade; it means the two files do
+#   not describe the same thing. KEPT 0 with a non-empty REMOVED list aborts. This is the
+#   general net behind the HOSTN check — the failure mode is silent and total, so it fires in
+#   --dry-run as well, putting the warning in the report itself.
+#
 # Root Required to Write
 #   Installing over the target requires root. --dry-run deliberately does not, so the
 #   change report can be previewed by anyone.
@@ -148,6 +164,30 @@ done
 [[ -f "$TEMPLATE" ]] || { echo "Error: template not found: $TEMPLATE" >&2; exit 1; }
 [[ -f "$TARGET"   ]] || { echo "Error: target not found: $TARGET"     >&2; exit 1; }
 
+# ── Guard: unsubstituted HOSTN placeholders ──────────────────────────────────────────────────
+#
+# host.conf.template ships HOSTN_ placeholders; a live host conf uses HOST1_ / HOST2_. Key
+# matching below is literal, so merging the raw template against a real host conf classifies
+# EVERY existing key as deprecated and every template key as new — KEPT 0, and the install
+# would drop every credential in the file. api/setup.php substitutes before writing; a direct
+# invocation has no such step. Refused rather than auto-substituted: this rewrites the file
+# that holds every secret on the host, and guessing the slot is not worth being wrong about.
+if grep -q 'HOSTN' "$TEMPLATE" 2>/dev/null; then
+    _slot=$(grep -oEm1 '^[[:space:]]*(HOST[0-9]+)_' "$TARGET" 2>/dev/null | grep -oE 'HOST[0-9]+')
+    if [[ -n "$_slot" ]]; then
+        _lower=$(echo "$_slot" | tr '[:upper:]' '[:lower:]')
+        echo "Error: template still contains HOSTN placeholders, but the target uses ${_slot}_." >&2
+        echo "       Merging as-is would classify all ${_slot}_ keys as deprecated and remove" >&2
+        echo "       them — including every credential. Substitute the slot first:" >&2
+        echo "" >&2
+        echo "         sed -e 's/HOSTN/${_slot}/g' -e 's/hostn/${_lower}/g' \\" >&2
+        echo "             $TEMPLATE > /tmp/${_lower}.conf.template" >&2
+        echo "         $0 --template /tmp/${_lower}.conf.template --target $TARGET --dry-run" >&2
+        echo "" >&2
+        exit 1
+    fi
+fi
+
 # ── Parse target → KEY → full definition block ───────────────────────────────────────────────
 
 declare -A HOST_MAP   # KEY → complete definition line(s) from user's conf
@@ -268,6 +308,24 @@ fi
 
 echo "  KEPT    ${#KEPT[@]} existing vars — your values preserved"
 
+# ── Guard: total mismatch ────────────────────────────────────────────────────────────────────
+#
+# Keeping nothing from a populated conf is never a real upgrade — it is the signature of the
+# two files not describing the same thing (wrong template, wrong target, unsubstituted
+# placeholders). The HOSTN check above catches the known cause; this catches the rest, because
+# the failure mode is silent and total: every value in the file is replaced by a default.
+# Deliberately fires in --dry-run too, so the report itself carries the warning.
+if [[ ${#KEPT[@]} -eq 0 && ${#REMOVED[@]} -gt 0 ]]; then
+    echo "────────────────────────────────────────────────────────────────────────"
+    echo ""
+    echo "Error: refusing — this would keep NOTHING and remove all ${#REMOVED[@]} existing keys." >&2
+    echo "       A genuine upgrade preserves values; keeping zero means the template and the" >&2
+    echo "       target do not describe the same conf. Check that --template matches --target" >&2
+    echo "       and that any HOSTN placeholders were substituted for this host's slot." >&2
+    echo "" >&2
+    exit 1
+fi
+
 if [[ ${#ADDED[@]} -eq 0 && ${#REMOVED[@]} -eq 0 ]]; then
     echo "  Already up to date — no changes needed."
     echo "────────────────────────────────────────────────────────────────────────"

Deployment/host.conf.template

@@ -580,7 +580,7 @@
     HOSTN_OLLAMA_URL=""                          # e.g. http://localhost:11434 — empty if no local Ollama
     HOSTN_OLLAMA_CONTAINER="Ollama"              # for watchdog / restart lists
     HOSTN_OLLAMA_GPU_UUID=""                     # pins Ollama to one card on multi-GPU hosts
-    HOSTN_OLLAMA_MODEL="qwen2.5-coder:14b"       # generation
+    HOSTN_OLLAMA_MODEL="hf.co/unsloth/Qwen3-14B-GGUF:IQ4_XS"  # generation — must fully offload; see README-AI.md
     HOSTN_OLLAMA_EMBED_MODEL="nomic-embed-text"  # embeddings — the generation model cannot embed
 
 # ━━━ Authelia ━━━

Notes_AI-Design.md

@@ -1,26 +1,6 @@
 # Varaverk AI Integration — Design Notes
 
-**Status: retrieval is built; integration is not.** As of 2026-08-02 the `AI_*` and
+**Status
-`HOST*_OLLAMA_*` variables exist in both confs and both templates, and `AI/` holds a working
-index and query path — see the RAG section at the end of this document and `AI/README-AI.md`.
-
-Everything else below remains design only. **No Varaverk script consults AI.** Every
-`AI_ASSIST_*` toggle is false, `AI_CONF_WRITE_ENABLED` is false with an empty whitelist, and
-host resolution across the mesh is specified but not implemented. Originally captured
-2026-08-01 so the reasoning survives.
-
-Ollama itself *is* installed, tuned and verified on HOST1 — `qwen2.5-coder:14b` for
-generation, `nomic-embed-text` for embeddings, 16k context, pinned to the RTX 3080. See
-Hardware Budget for measured numbers. That is the substrate, not the integration.
-
-**Origin:** the RTX 3080 was freed when the Windows gaming VM was retired. It is bound to the
-`nvidia` driver, not `vfio` — not reserved for passthrough, so there is no VM contention to
-design around. Two goals at once: somewhere to learn local LLMs, and something Varaverk can
-genuinely use.
-
-**Build order** — deliberately lowest-risk first. Each stage must be boring before the next
-one starts:
-
 1. Chat assistant / settings helper / onboarding assistant — a wrong answer costs nothing
 2. Watchdog and discovery context — a wrong answer costs a bad suggestion, still gated
 3. Cleanup and sync decision aid — closest to destructive, last to be trusted