diff --git a/docs/superpowers/evidence/website-force-recreate-2026-08-29.json b/docs/superpowers/evidence/website-force-recreate-2026-08-29.json
new file mode 100644
index 00000000..edbead20
--- /dev/null
+++ b/docs/superpowers/evidence/website-force-recreate-2026-08-29.json
@@ -0,0 +1,42 @@
+{
+  "intent": "Make the website deploy apply Caddyfile changes and survive a replaced dist/ directory, instead of reporting a healthy container that serves 404s.",
+  "scope": [
+    "linux_configuration/features/lib/website_build.sh (start_website and wait_for_website only)",
+    "Non-goal: no change to the Gitea stack, the firewall step, the build, or the edge reload"
+  ],
+  "changes": [
+    "start_website now runs docker compose up -d --force-recreate; compose cannot detect a bind-mounted Caddyfile's contents changing, nor a bind mount left on an unlinked inode, so it reported 'Running' and applied neither",
+    "wait_for_website records the actual HTTP status instead of using curl -sf, so a failure distinguishes 'no response' from 'answered 404' -- different bugs that had one message",
+    "The 404 message names the two real causes: a stale /srv bind mount, or an empty dist/"
+  ],
+  "verification": [
+    {
+      "command": "observed live after setup_personal_website.sh",
+      "result": "fail",
+      "evidence": "Container 'Up 6 hours', Caddy logs healthy, but docker exec ls /srv returned empty and curl 127.0.0.1:8088/ returned 404; dist/ had been replaced with a new inode (84695030) after container start"
+    },
+    {
+      "command": "rm -rf dist && pnpm build, then curl",
+      "result": "fail",
+      "evidence": "Reproduced deterministically: inode 84695030 -> 84695031, container /srv empty, curl -> 404"
+    },
+    {
+      "command": "source website_build.sh && start_website (patched), against that broken state",
+      "result": "pass",
+      "evidence": "Container Recreated; /srv lists assets blog cv.pdf index.html; / -> 200, /blog/hello-world -> 200 with title 'Hello, world — Krzysztof Rudnicki'"
+    },
+    {
+      "command": "shellcheck + bash -n linux_configuration/features/lib/website_build.sh",
+      "result": "pass",
+      "evidence": "no findings"
+    }
+  ],
+  "risks": [
+    "Every deploy now restarts the static container, so there is a sub-second window where the loopback port refuses connections; the edge Caddy retries and the wait loop tolerates it",
+    "--force-recreate discards any state inside the container; this service is a read-only static file server and holds none"
+  ],
+  "rollback": [
+    "Drop --force-recreate and restore the curl -sf wait loop",
+    "After rollback, verify a Caddyfile edit actually takes effect by curling a prerendered blog post and checking its <title>, since that is what silently regresses"
+  ]
+}
diff --git a/linux_configuration/features/lib/website_build.sh b/linux_configuration/features/lib/website_build.sh
index 8cd861e9..66e435a0 100755
--- a/linux_configuration/features/lib/website_build.sh
+++ b/linux_configuration/features/lib/website_build.sh
@@ -76,10 +76,17 @@ EOF
 }
 
 # --- Phase 8: start the website container -----------------------------------
+# Last HTTP status seen by wait_for_website, so a failure can say whether the
+# server was unreachable or merely answering wrongly. Those are different bugs
+# and the old message named only the first one.
+WEBSITE_LAST_CODE=""
+
 wait_for_website() {
 	local _
 	for _ in $(seq 1 30); do
-		if curl -sf -o /dev/null "http://127.0.0.1:${WEBSITE_PORT}/"; then
+		WEBSITE_LAST_CODE="$(curl -s -o /dev/null -w '%{http_code}' \
+			--max-time 5 "http://127.0.0.1:${WEBSITE_PORT}/" || echo "000")"
+		if [[ "$WEBSITE_LAST_CODE" == "200" ]]; then
 			return 0
 		fi
 		sleep 1
@@ -89,11 +96,30 @@ wait_for_website() {
 
 start_website() {
 	log_info "Starting the personal-website container…"
-	docker compose -f "$WEBSITE_COMPOSE" up -d
+	# --force-recreate is required, not tidiness. Two failure modes need it,
+	# and compose cannot see either one:
+	#
+	#   1. The Caddyfile is a BIND MOUNT, so its contents are not part of the
+	#      container's config hash. Editing it changes nothing compose compares,
+	#      it reports "Running", and the new directives never load. A blog post
+	#      served as the landing page is exactly this bug.
+	#   2. A bind mount is resolved once, at container start. Anything that
+	#      replaces dist/ rather than emptying it (rm -rf, a git clean, moving
+	#      the checkout) leaves the container mounted on the old, unlinked
+	#      inode -- /srv reads as empty and every request 404s while the
+	#      container still looks healthy.
+	#
+	# Recreating is under a second for a static file server, so it is cheaper
+	# than either failure being diagnosed by hand.
+	docker compose -f "$WEBSITE_COMPOSE" up -d --force-recreate
 	if wait_for_website; then
 		log_ok "Website answering on http://127.0.0.1:${WEBSITE_PORT}/."
 	else
-		die "Website did not become reachable on 127.0.0.1:${WEBSITE_PORT}."
+		if [[ "$WEBSITE_LAST_CODE" == "000" ]]; then
+			die "Website unreachable on 127.0.0.1:${WEBSITE_PORT} (no response)."
+		fi
+		die "Website on 127.0.0.1:${WEBSITE_PORT} answered HTTP ${WEBSITE_LAST_CODE}, not 200. \
+A 404 here usually means the container's /srv bind mount is stale or ${WEBSITE_SRC}/dist is empty."
 	fi
 }