diff --git a/scripts/devnet.sh b/scripts/devnet.sh
index c89acfed7..2d1c24654 100755
--- a/scripts/devnet.sh
+++ b/scripts/devnet.sh
@@ -101,10 +101,36 @@ ensure_built() {
 
 start_hardhat() {
   local pidfile="$DEVNET_DIR/hardhat.pid"
+  local marker="$DEVNET_DIR/hardhat/deployed"
+  local artifacts_dir="$REPO_ROOT/packages/evm-module/artifacts/contracts"
 
   if [ -f "$pidfile" ] && kill -0 "$(cat "$pidfile")" 2>/dev/null; then
-    log "Hardhat node already running (PID $(cat "$pidfile"))"
-    return 0
+    # Stale-bytecode guard. When contracts are recompiled while hardhat
+    # keeps running, the on-chain bytecode lags the artifacts on disk.
+    # The next `devnet start` short-circuits hardhat reuse here, then
+    # `deploy_contracts` skips on the still-present marker, so daemons
+    # connect to addresses whose code is older than the source. View
+    # methods added since the live deploy revert with "function selector
+    # not recognized" (e.g. CSS.getNodeStakeV10 in the staking loop).
+    # Detect the mismatch by comparing artifact mtimes to the marker;
+    # if any artifact is newer, kill hardhat so the fresh-start path
+    # below clears the marker + deployments and `deploy_contracts`
+    # re-deploys onto a fresh chain.
+    if [ -f "$marker" ] && [ -d "$artifacts_dir" ] \
+        && [ -n "$(find "$artifacts_dir" -name '*.json' -newer "$marker" -print -quit 2>/dev/null)" ]; then
+      local stale_pid
+      stale_pid=$(cat "$pidfile")
+      log "Hardhat (PID $stale_pid) holds outdated contracts (artifacts newer than deploy marker) — restarting for fresh deploy"
+      kill "$stale_pid" 2>/dev/null || true
+      for _ in 1 2 3 4 5; do
+        kill -0 "$stale_pid" 2>/dev/null || break
+        sleep 1
+      done
+      rm -f "$pidfile"
+    else
+      log "Hardhat node already running (PID $(cat "$pidfile"))"
+      return 0
+    fi
   fi
 
   log "Starting Hardhat node on port $HARDHAT_PORT..."
@@ -739,7 +765,7 @@ cmd_start() {
       const opSigners = new Array(n).fill(null);
       const nodeRoles = new Array(n).fill('edge');
       // Codex round 4 on PR #368: read config.json FIRST and INDEPENDENTLY
-      // of wallets.json. The previous loop did wallets first and `continue`d
+      // of wallets.json. The previous loop did wallets first and \`continue\`d
       // on parse failure BEFORE reading config, so an intended core whose
       // wallets.json was malformed silently kept the 'edge' default,
       // dropped out of coreIdxs, and the lostCores guard never fired.
@@ -865,7 +891,7 @@ cmd_start() {
         // this code path exists to prevent.
         //
         // Codex round 2 on PR #368: re-read 'pending' before EVERY tx
-        // (not once per node loop). The previous code did `nonce++` at
+        // (not once per node loop). The previous code did \`nonce++\` at
         // call site so a failed approve advanced the local counter even
         // though the chain nonce did not, leaving createConviction +
         // updateAsk to send with an inflated nonce that would either
@@ -878,7 +904,7 @@ cmd_start() {
           return tx.wait();
         };
         // Codex round 4 on PR #368: skip createConviction if the daemon
-        // already opened a position. PR 366 wired `EVMChainAdapter.ensureProfile()`
+        // already opened a position. PR 366 wired \`EVMChainAdapter.ensureProfile()\`
         // to open a default 50k V10 conviction during agent startup, so by
         // the time this script reaches the staking loop the daemon has
         // (usually) already staked. Running createConviction again would
@@ -895,8 +921,8 @@ cmd_start() {
         // updateAsk is INDEPENDENT of stake state, so we still attempt
         // it below regardless of the probe outcome.
         //
-        // Codex round 7 on PR #368: probe `getNodeStakeV10(idId)` instead
-        // of NFT `balanceOf(opSigner.address)`. The op wallet could hold
+        // Codex round 7 on PR #368: probe \`getNodeStakeV10(idId)\` instead
+        // of NFT \`balanceOf(opSigner.address)\`. The op wallet could hold
         // positions for OTHER identities (not in our devnet today, but
         // a wallet-scoped probe is the wrong semantic check); we want
         // to know whether THIS identity is staked.