diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 1c12c23..4181a82 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -25,6 +25,7 @@ permissions:
 jobs:
   test:
     runs-on: ubuntu-latest
+    timeout-minutes: 25
     outputs:
       passed: ${{ steps.result.outputs.passed }}
       coverage: ${{ steps.result.outputs.coverage }}
@@ -58,6 +59,7 @@ jobs:
   mcdc:
     # 100% MC/DC (condition coverage) enforced on every Tier-1 driver harness.
     runs-on: ubuntu-latest
+    timeout-minutes: 15
     steps:
       - uses: actions/checkout@v4
       - name: install gcc-14
@@ -78,6 +80,7 @@ jobs:
     # Smoke-run the sim micro-benchmarks and emit the ns/op trend JSON. Regression
     # gating against stored history is future work (no flare-edge pattern to copy).
     runs-on: ubuntu-latest
+    timeout-minutes: 20
     steps:
       - uses: actions/checkout@v4
       - name: run sim benchmarks
@@ -99,6 +102,7 @@ jobs:
     # fail-closed sha), and build the A/B disk image (unprivileged mkfs -d).
     # Booting needs a zImage and therefore lives in kernel-build's smoke step.
     runs-on: ubuntu-latest
+    timeout-minutes: 20
     steps:
       - uses: actions/checkout@v4
       - name: shellcheck qemu scripts
@@ -118,6 +122,7 @@ jobs:
   patches-apply:
     # The RV1106 series must apply cleanly onto pristine linux-6.18.46.
     runs-on: ubuntu-latest
+    timeout-minutes: 20
     steps:
       - uses: actions/checkout@v4
       - name: cache pristine kernel tarball
@@ -125,11 +130,8 @@ jobs:
         with:
           path: ~/linux-6.18.46.tar.xz
           key: linux-6.18.46-tarball
-      - name: fetch + verify pristine
-        run: |
-          [ -f ~/linux-6.18.46.tar.xz ] || \
-            curl -fSL https://cdn.kernel.org/pub/linux/kernel/v6.x/linux-6.18.46.tar.xz -o ~/linux-6.18.46.tar.xz
-          echo "$(cat build/linux-6.18.46.tar.xz.sha256)  $HOME/linux-6.18.46.tar.xz" | sha256sum -c -
+      - name: fetch + verify pristine (shared fail-closed fetcher)
+        run: bash build/fetch-kernel-tarball.sh "$HOME/linux-6.18.46.tar.xz"
       - name: apply the series in order
         run: |
           tar -C /tmp -xf ~/linux-6.18.46.tar.xz
@@ -148,6 +150,7 @@ jobs:
   prune-artifacts:
     if: github.event_name == 'workflow_dispatch'
     runs-on: ubuntu-latest
+    timeout-minutes: 10
     permissions:
       actions: write
     steps:
@@ -178,13 +181,14 @@ jobs:
     if: github.event_name == 'workflow_dispatch'
     needs: [prune-artifacts]
     runs-on: ubuntu-latest
+    timeout-minutes: 60
     steps:
       - uses: actions/checkout@v4
       - name: install cross toolchain + kernel build deps + qemu
         run: |
           sudo apt-get update -qq
           sudo apt-get install -y -qq gcc-arm-linux-gnueabihf qemu-system-arm \
-            cpio bc bison flex libssl-dev
+            cpio bc bison flex libssl-dev ccache
       - name: provision `python` (SDK quirk — build calls bare python)
         run: |
           mkdir -p "$RUNNER_TEMP/bin"
@@ -195,22 +199,32 @@ jobs:
         with:
           path: ~/linux-6.18.46.tar.xz
           key: linux-6.18.46-tarball
+      # Ephemeral runners rebuild the whole tree every dispatch (~9 min of
+      # compile); ccache recovers most of it for an unchanged/lightly-changed
+      # series. Keyed on the config + patches so a real change misses cleanly.
+      - name: cache ccache
+        uses: actions/cache@v4
+        with:
+          path: ~/.ccache
+          key: kbuild-ccache-${{ hashFiles('build/warden_defconfig', 'patches/*.patch') }}
+          restore-keys: kbuild-ccache-
       - name: build zImage + rv1106-warden.dtb
         env:
           # WORK must be OUTSIDE the repo checkout: build-kernel.sh applies the patch
           # series with `git apply`, which silently ignores out-of-subdir paths when
           # run inside another git repo (issue #1). $RUNNER_TEMP is outside the checkout.
           WORK: ${{ runner.temp }}/kbuild-out
-          # Reuse the cached tarball when present (build-kernel.sh still verifies
-          # the sha256 pin fail-closed either way; it downloads if the file is absent).
-          KERNEL_TARBALL: ~/linux-6.18.46.tar.xz
           # The kernel is freestanding; the generic arm cross toolchain links it.
           CROSS_COMPILE: arm-linux-gnueabihf-
+          WARDEN_CCACHE: 1
+          CCACHE_DIR: /home/runner/.ccache
+        # KERNEL_TARBALL is exported from the SHELL so $HOME expands — a literal
+        # `~` in a YAML env: value is never tilde-expanded and broke every
+        # dispatch until caught in review.
         run: |
-          [ -f ~/linux-6.18.46.tar.xz ] || \
-            curl -fSL https://cdn.kernel.org/pub/linux/kernel/v6.x/linux-6.18.46.tar.xz \
-              -o ~/linux-6.18.46.tar.xz
+          export KERNEL_TARBALL="$HOME/linux-6.18.46.tar.xz"
           bash build/build-kernel.sh
+          ccache -s | head -4
       # Boot smoke under QEMU: the zImage this job just built must reach the
       # initramfs sentinel on -M virt (verified 2026-08-29: the canonical
       # config boots virt as-is). FAIL-CLOSED on a missing qemu-system-arm.
@@ -239,6 +253,7 @@ jobs:
     needs: [test]
     if: github.event_name == 'push' && github.ref == 'refs/heads/main'
     runs-on: ubuntu-latest
+    timeout-minutes: 15
     permissions:
       contents: write
     steps:
diff --git a/.gitignore b/.gitignore
index cb4c663..8493652 100644
--- a/.gitignore
+++ b/.gitignore
@@ -14,7 +14,7 @@ target/
 __pycache__/
 *.pyc
 
-# local tooling state (code-review harness cross-session memory, etc.)
+# local tooling state
 .claude/
 
 # driver MC/DC harness build dirs
diff --git a/README.md b/README.md
index d51b25c..e39d0ca 100644
--- a/README.md
+++ b/README.md
@@ -121,7 +121,9 @@ Evaluated against the stack philosophy — **openness, hardness, modernness**:
 ## Relationship to flare-edge
 
 flare-edge (WardenOS: the LVGL UI + the `flared` daemon) is the product; warden-sdk
-is what builds and tests it. During bootstrap, flare-edge consumes warden-sdk piece
+is what builds and tests it. flare-edge is BlueFlare's private companion repo —
+not publicly available — so flare-edge issue references and checkout paths in
+this repo's docs are context, not reachable links. During bootstrap, flare-edge consumes warden-sdk piece
 by piece: first the simulator (as a dev/test dependency), later the image build.
 No flare-edge code moves here — only the SDK/build/sim/driver-seam layer.
 
diff --git a/build/build-kernel.sh b/build/build-kernel.sh
index b346ef7..f5db1e9 100755
--- a/build/build-kernel.sh
+++ b/build/build-kernel.sh
@@ -14,7 +14,9 @@
 #   JOBS             parallel make jobs (default: nproc)
 #   WARDEN_KCONFIG_FRAGMENT
 #                    optional kconfig fragment merged onto warden_defconfig
-#                    (qemu/configs/virt.fragment builds the QEMU -M virt variant)
+#                    (qemu/configs/virt.fragment builds the QEMU -M virt variant);
+#                    every fragment option is verified to have taken effect
+#   WARDEN_CCACHE=1  compile through ccache (CI caches ~/.ccache)
 #
 # Requires: `python` (not python3) on PATH — the SDK quirk; the CI runner provides
 # a project-local venv. Builds are SERIAL on the shared SDK box — never run two.
@@ -24,9 +26,8 @@ KVER=6.18.46
 HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"     # build/
 REPO="$(cd "$HERE/.." && pwd)"
 PATCHES="$REPO/patches"
-SHA_FILE="$HERE/linux-$KVER.tar.xz.sha256"
 JOBS="${JOBS:-$(nproc)}"
-URL="https://cdn.kernel.org/pub/linux/kernel/v6.x/linux-$KVER.tar.xz"
+# The tarball URL + sha256 pin live in fetch-kernel-tarball.sh (shared with CI).
 
 # A caller-provided WORK (e.g. CI's ${{ github.workspace }}/kbuild-out, from which
 # artifacts are uploaded) is left intact; a scratch dir we mktemp'd here is our own
@@ -47,21 +48,9 @@ command -v python >/dev/null || { echo "need 'python' (not python3) on PATH —
 # 1. obtain + verify the pristine tarball
 mkdir -p "$WORK"
 TB="${KERNEL_TARBALL:-$WORK/linux-$KVER.tar.xz}"
-if [ ! -f "$TB" ]; then
-  log "downloading $URL"
-  curl -fSL "$URL" -o "$TB"
-fi
-# Fail closed: a missing pin (e.g. forgotten on a KVER bump) or a KERNEL_TARBALL
-# pointed at an arbitrary file must refuse to build, never silently skip the check
-# — the pristine tarball is the ONLY external input and integrity is the whole point.
-[ -f "$SHA_FILE" ] || {
-  echo "FATAL: no pinned sha256 for linux-$KVER (expected $SHA_FILE) — refusing to build from an unverified tarball" >&2
-  exit 1
-}
-want="$(cat "$SHA_FILE")"
-got="$(sha256sum "$TB" | awk '{print $1}')"
-[ "$want" = "$got" ] || { echo "tarball sha256 mismatch: want $want got $got" >&2; exit 1; }
-log "tarball sha256 verified"
+# Fetch + fail-closed sha256 verification live in ONE place shared with CI
+# (a missing pin or a mismatch always refuses to build).
+bash "$HERE/fetch-kernel-tarball.sh" "$TB"
 
 # 2. extract pristine
 SRC="$WORK/linux-$KVER"
@@ -124,9 +113,41 @@ command -v "${CROSS_COMPILE}gcc" >/dev/null \
   || { echo "cross toolchain ${CROSS_COMPILE}gcc not on PATH (set SDK_TC, or CROSS_COMPILE to one that is)" >&2; exit 1; }
 make -C "$SRC" ARCH=arm CROSS_COMPILE="$CROSS_COMPILE" olddefconfig >/dev/null
 
+# Fragment took-effect assertion: merge_config -m only pastes text, and
+# olddefconfig silently resolves any symbol whose dependencies are unmet —
+# a fragment option could be dropped without a word. Verify every explicit
+# request in the fragment survived into the final .config; fail loud if not.
+if [ -n "${WARDEN_KCONFIG_FRAGMENT:-}" ]; then
+  frag_fail=0
+  while IFS= read -r line; do
+    case "$line" in
+      CONFIG_*=*)
+        grep -qxF "$line" "$SRC/.config" || {
+          echo "FATAL: fragment option '$line' did not take effect (unmet Kconfig dependency?)" >&2
+          frag_fail=1
+        } ;;
+      "# CONFIG_"*" is not set")
+        opt="${line#\# }"; opt="${opt% is not set}"
+        grep -qE "^$opt=" "$SRC/.config" && {
+          echo "FATAL: fragment disabled '$opt' but it is set in the final .config" >&2
+          frag_fail=1
+        } ;;
+    esac
+  done < "$FRAG"
+  [ "$frag_fail" = 0 ] || exit 1
+  log "fragment options verified in final .config"
+fi
+
+# Optional ccache (CI: cache ~/.ccache across dispatches; harmless if unset).
+KCC="${CROSS_COMPILE}gcc"
+if [ "${WARDEN_CCACHE:-0}" = 1 ]; then
+  command -v ccache >/dev/null || { echo "FATAL: WARDEN_CCACHE=1 but ccache not installed" >&2; exit 1; }
+  KCC="ccache ${CROSS_COMPILE}gcc"
+fi
+
 # 5. build zImage + the board dtb
 log "building zImage + rv1106-warden.dtb (-j$JOBS)"
-make -C "$SRC" ARCH=arm CROSS_COMPILE="$CROSS_COMPILE" -j"$JOBS" \
+make -C "$SRC" ARCH=arm CROSS_COMPILE="$CROSS_COMPILE" CC="$KCC" -j"$JOBS" \
   zImage rockchip/rv1106-warden.dtb
 
 Z="$SRC/arch/arm/boot/zImage"
diff --git a/build/fetch-kernel-tarball.sh b/build/fetch-kernel-tarball.sh
new file mode 100644
index 0000000..ffc996a
--- /dev/null
+++ b/build/fetch-kernel-tarball.sh
@@ -0,0 +1,28 @@
+#!/usr/bin/env bash
+# Fetch (with retries) and sha256-verify the pristine kernel tarball into $1.
+# Single source of truth for the URL + verification used by build-kernel.sh
+# and both CI jobs (patches-apply, kernel-build) — a KVER bump edits this file
+# and build-kernel.sh only. FAILS CLOSED: a missing pin refuses to proceed.
+#
+# Usage: fetch-kernel-tarball.sh <destination-path>
+set -euo pipefail
+
+KVER=6.18.46
+HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"     # build/
+SHA_FILE="$HERE/linux-$KVER.tar.xz.sha256"
+URL="https://cdn.kernel.org/pub/linux/kernel/v6.x/linux-$KVER.tar.xz"
+
+TB="${1:?usage: fetch-kernel-tarball.sh <destination-path>}"
+
+if [ ! -f "$TB" ]; then
+  echo "== downloading $URL"
+  curl --retry 3 --retry-delay 5 -fSL "$URL" -o "$TB"
+fi
+[ -f "$SHA_FILE" ] || {
+  echo "FATAL: no pinned sha256 for linux-$KVER (expected $SHA_FILE) — refusing an unverified tarball" >&2
+  exit 1
+}
+want="$(cat "$SHA_FILE")"
+got="$(sha256sum "$TB" | awk '{print $1}')"
+[ "$want" = "$got" ] || { echo "FATAL: tarball sha256 mismatch: want $want got $got" >&2; exit 1; }
+echo "== tarball sha256 verified: $TB"
diff --git a/docs/architecture.md b/docs/architecture.md
index ffb3ebf..812773d 100644
--- a/docs/architecture.md
+++ b/docs/architecture.md
@@ -65,7 +65,7 @@ supervisor logic runs in CI with no panel.
   the boot-mode register's survives-warm-reset / cleared-by-POR behaviour (the MaskRom
   recovery maneuver). The matching firmware-side `Bus` seam on flared's `devmem` — so
   the shipped ladder can be asserted to poke the confirmed offset, never the wrong-SoC
-  one — lands when flare-edge consumes warden-sdk (§7 item 3, maintainer-gated), not yet on
+  one — lands when flare-edge consumes warden-sdk (§8 item 3, maintainer-gated), not yet on
   flare-edge `main`.
 - **`modbus` — RS-485 device end.** Done. `ModbusSlave`: a byte-in/byte-out RTU slave
   (CRC16 byte-identical to the master, FC 0x01–0x06/0x0F/0x10/0x11, exception replies,
diff --git a/docs/ci-cd.md b/docs/ci-cd.md
index 1389666..5e7794e 100644
--- a/docs/ci-cd.md
+++ b/docs/ci-cd.md
@@ -2,8 +2,8 @@
 
 `.github/workflows/ci.yml` — every job runs on GitHub-hosted `ubuntu-latest`.
 No self-hosted runner is (or may be) reachable from this repo's workflows:
-the repo is public, and a fork PR that gets one approved run could otherwise
-execute code on private infrastructure (ADR-0007).
+the repo is going public, and a fork PR that gets one approved run could
+otherwise execute code on private infrastructure (ADR-0007).
 
 ## Jobs
 
diff --git a/docs/decisions/0003-standalone-repo.md b/docs/decisions/0003-standalone-repo.md
index 0b15fd2..b08f93c 100644
--- a/docs/decisions/0003-standalone-repo.md
+++ b/docs/decisions/0003-standalone-repo.md
@@ -1,6 +1,8 @@
 # ADR 0003 — warden-sdk is a standalone repo
 
-**Status:** Accepted (2026-08-25).
+**Status:** Accepted (2026-08-25). Repo-visibility half superseded by ADR-0007
+(2026-08-30) — warden-sdk went public; the "private for now" consequence below
+no longer holds. Original decision kept for the record.
 
 ## Context
 Our real SDK changes lived as uncommitted edits in a 2GB opaque vendor fork, with
@@ -10,7 +12,7 @@ no CI, tests, or versioning of their own. The SDK requirement (future-features-2
 ## Decision
 A **private** `bfe-noah/warden-sdk` GitHub repo, standalone from day one with its own
 CI/versioning. Work lands on a `bringup` branch; the first commit to `main` is gated
-on a passing code-review-harness run, green CI, and the maintainer's fresh explicit go-ahead.
+on a passing review run, green CI, and the maintainer's fresh explicit go-ahead.
 
 ## Consequences
 - flare-edge consumes warden-sdk later (flared depending on `warden-sim`, drivers
diff --git a/kernel/rv1106-enablement/build-m2.sh b/kernel/rv1106-enablement/build-m2.sh
index 9a651dd..c8b5f90 100755
--- a/kernel/rv1106-enablement/build-m2.sh
+++ b/kernel/rv1106-enablement/build-m2.sh
@@ -8,7 +8,7 @@
 # Defaults match this workspace.
 set -euo pipefail
 
-FE="${FE:-<flare-edge>}"
+FE="${FE:?set FE to a flare-edge checkout path}"
 KTREE="${KTREE:-$FE/research/linux-6.18.46}"
 SDK_TC="${SDK_TC:-$FE/sdk/tools/linux/toolchain/arm-rockchip830-linux-uclibcgnueabihf/bin}"
 HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
diff --git a/kernel/rv1106-enablement/mailbox/scr1-echo/Makefile b/kernel/rv1106-enablement/mailbox/scr1-echo/Makefile
index 5008b9f..54678a7 100644
--- a/kernel/rv1106-enablement/mailbox/scr1-echo/Makefile
+++ b/kernel/rv1106-enablement/mailbox/scr1-echo/Makefile
@@ -1,7 +1,8 @@
 # HPMCU mailbox-echo firmware — bare-metal RV32IMC for the RV1106 SCR1 core.
 # Same xPack riscv-none-embed-gcc 10.2.0 + flags as the watchdog firmware.
 
-XPACK ?= <flare-edge>/sdk/sysdrv/source/mcu/prebuilts/gcc/linux-x86/riscv64/xpack-riscv-none-embed-gcc-10.2.0-1.2/bin
+# Point XPACK at the xpack riscv toolchain bin/ inside a flare-edge SDK checkout.
+XPACK ?= $(error set XPACK to the xpack-riscv-none-embed-gcc bin/ directory)
 CROSS ?= $(XPACK)/riscv-none-embed-
 
 CC      = $(CROSS)gcc
diff --git a/kernel/rv1106-enablement/rga/PORT-PLAN.md b/kernel/rv1106-enablement/rga/PORT-PLAN.md
index 79f1cea..6e1c01b 100644
--- a/kernel/rv1106-enablement/rga/PORT-PLAN.md
+++ b/kernel/rv1106-enablement/rga/PORT-PLAN.md
@@ -295,7 +295,8 @@ implied by this kernel port.
   canvas + scanout mirror at 720×720×4B ≈ 2 MiB each); carry the same size unless a
   future accounting shows it's tight.
 - **Driver-parity table** (`../DRIVER-PARITY.md`) should move `RGA 2D (rga2)` from [ ] to
-  [wip]/[x] as these steps land, same convention as every other M-milestone row.
+  [wip]/[x] as these steps land, same convention as every other M-milestone row
+(the parity row already reads [x]).
 
 ## Sources
 
diff --git a/qemu/README.md b/qemu/README.md
index 37d81c8..02af94b 100644
--- a/qemu/README.md
+++ b/qemu/README.md
@@ -89,7 +89,8 @@ stage-2 init when present.
 - A serial port that is closed discards incoming bytes: hold ONE fd open
   across write and read when scripting the guest side of the RS485 bridge.
 - `highmem=off` and `-global virtio-mmio.force-legacy=false` are load-bearing
-  (32-bit ECAM reach; virtio-1-only gpu/input) — both live in run.sh.
+  (32-bit ECAM reach; virtio-1-only gpu/input) — both live ONLY in run.sh,
+  which every script (boot smoke included) delegates to.
 - Never pass `earlyprintk`: DEBUG_UART_PHYS is the RV1106's 0xff4c0000.
 
 ## Host requirements
diff --git a/qemu/lib.sh b/qemu/lib.sh
index cd8b2ed..4249709 100644
--- a/qemu/lib.sh
+++ b/qemu/lib.sh
@@ -18,7 +18,7 @@ qemu_get_busybox() {
   BB="${BUSYBOX:-$out/busybox-armv7l}"
   if [ ! -f "$BB" ]; then
     qemu_log "downloading $BB_URL"
-    curl -fSL "$BB_URL" -o "$BB"
+    curl --retry 3 --retry-delay 5 -fSL "$BB_URL" -o "$BB"
   fi
   [ -f "$sha_file" ] || {
     echo "FATAL: no pinned sha256 for busybox (expected $sha_file) — refusing to build from an unverified binary" >&2
diff --git a/qemu/mkimage.sh b/qemu/mkimage.sh
index 96036ba..7d38e98 100755
--- a/qemu/mkimage.sh
+++ b/qemu/mkimage.sh
@@ -30,7 +30,17 @@ FW_VERSION="0.0.1"
 while [ $# -gt 0 ]; do
   case "$1" in
     --portal-url) PORTAL_URL="${2:?--portal-url needs a value}"; shift 2 ;;
-    --state)      STATE_KV+=("${2:?--state needs KEY=VALUE}"); shift 2 ;;
+    --state)
+      case "${2:?--state needs KEY=VALUE}" in
+        *=*) ;;
+        *) echo "FATAL: --state needs KEY=VALUE, got '$2'" >&2; exit 1 ;;
+      esac
+      case "${2%%=*}" in
+        *[!A-Za-z0-9_.]*|'')
+          echo "FATAL: --state key '${2%%=*}' must match [A-Za-z0-9_.]+ (it becomes a filename)" >&2
+          exit 1 ;;
+      esac
+      STATE_KV+=("$2"); shift 2 ;;
     --fw-version) FW_VERSION="${2:?--fw-version needs a value}"; shift 2 ;;
     *) echo "FATAL: unknown argument '$1' (usage: mkimage.sh [--portal-url URL] [--state KEY=VALUE]... [--fw-version V])" >&2; exit 1 ;;
   esac
@@ -83,7 +93,11 @@ place_partition() {
     rootfs_a|rootfs_b) stage="$ROOT" ;;
     userdata)          stage="$UDATA" ;;
     oem_a|oem_b)       stage="$SCRATCH/empty" ;;
-    *)                 stage="" ;;   # boot-chain partition: left zeroed
+    # Boot-chain partitions the VM never reads: present at the right offsets,
+    # left zeroed. Enumerated (not a wildcard) so a typo'd name in
+    # blkdevparts.conf fails HERE, not as a confusing mount error at boot.
+    env|idblock|uboot|misc|boot_a|boot_b|recovery) stage="" ;;
+    *) echo "FATAL: unknown partition name '$name' in blkdevparts.conf" >&2; exit 1 ;;
   esac
   # dd in 4K blocks — every offset in the canonical layout is 4K-aligned;
   # assert rather than assume, a misaligned write would corrupt a neighbor.
@@ -91,7 +105,8 @@ place_partition() {
     echo "FATAL: partition $name not 4K-aligned (off=$off size=$size)" >&2
     exit 1
   fi
-  DISK_END=$((off + size))
+  # Max, not last: blkdevparts grammar permits explicit @offsets out of order.
+  [ $((off + size)) -gt "$DISK_END" ] && DISK_END=$((off + size))
   [ -z "$stage" ] && return 0
   local img="$SCRATCH/$name.img"
   mkfs_part "$stage" "$size" "$img"
diff --git a/qemu/rootfs/etc/rc b/qemu/rootfs/etc/rc
index f868fd5..fa2c335 100755
--- a/qemu/rootfs/etc/rc
+++ b/qemu/rootfs/etc/rc
@@ -1,37 +1,19 @@
 #!/bin/busybox sh
 # Stage-1 rc: sourced by /init (still PID 1, initramfs root) when a virtio
-# disk is present. Emulates U-Boot's slot choice — parse warden.slot= from the
-# cmdline, mount that rootfs, switch_root into it. This is an EMULATION of the
-# A/B selection outcome, not the BCB/bootcount mechanism itself.
+# disk is present. Emulates U-Boot's slot choice — mount the validated slot's
+# rootfs and switch_root into it. This is an EMULATION of the A/B selection
+# outcome, not the BCB/bootcount mechanism itself.
+#
+# Every guarded failure path `return`s to /init (valid in a sourced script;
+# /init then falls through to shell/poweroff). The final exec is the one
+# unguardable step: if switch_root itself fails to launch, the shell — PID 1 —
+# exits and the kernel panics; the applet-existence check below catches the
+# only preventable variant of that.
 
-# /dev/block/by-name/<PARTNAME> symlinks: the contract flare-edge slotctl.rs
-# relies on. blkdevparts= gives every vda partition a PARTNAME in sysfs.
-mkdir -p /dev/block/by-name
-for uev in /sys/class/block/vda*/uevent; do
-    [ -f "$uev" ] || continue
-    partname=""
-    devname=""
-    while IFS='=' read -r k v; do
-        case "$k" in
-            PARTNAME) partname="$v" ;;
-            DEVNAME)  devname="$v" ;;
-        esac
-    done < "$uev"
-    [ -n "$partname" ] && [ -n "$devname" ] \
-        && ln -sf "/dev/$devname" "/dev/block/by-name/$partname"
-done
+. /etc/warden-lib.sh
 
-# Slot select: whole-token parse of warden.slot= (never a substring match).
-slot="_a"
-for tok in $(cat /proc/cmdline); do
-    case "$tok" in
-        warden.slot=*) slot="${tok#warden.slot=}" ;;
-    esac
-done
-case "$slot" in
-    _a|_b) ;;
-    *) echo "rc: bad warden.slot='$slot', falling back to _a"; slot="_a" ;;
-esac
+warden_populate_by_name
+slot="$(warden_slot)"
 
 root="/dev/block/by-name/rootfs${slot}"
 if [ ! -e "$root" ]; then
@@ -49,6 +31,11 @@ if [ ! -x /mnt/sbin/init ]; then
     umount /mnt
     return 0
 fi
+if ! command -v switch_root >/dev/null; then
+    echo "rc: busybox lacks switch_root — staying in initramfs"
+    umount /mnt
+    return 0
+fi
 
 echo "rc: switching root to rootfs${slot} ($root)"
 exec switch_root /mnt /sbin/init
diff --git a/qemu/rootfs/etc/warden-lib.sh b/qemu/rootfs/etc/warden-lib.sh
new file mode 100644
index 0000000..72801dc
--- /dev/null
+++ b/qemu/rootfs/etc/warden-lib.sh
@@ -0,0 +1,42 @@
+# shellcheck shell=sh
+# Shared helpers for the VM's stage-1 (/init + /etc/rc, initramfs) and stage-2
+# (/sbin/init, disk rootfs) boot scripts. Present in both filesystems because
+# both are staged from the same qemu/rootfs/ skeleton. ONE copy of each rule —
+# the slot-validation drift between two hand-copied parsers was a real
+# review finding.
+
+# Populate /dev/block/by-name/<PARTNAME> symlinks from sysfs uevents — the
+# contract flare-edge's slotctl.rs relies on. blkdevparts= gives every vda
+# partition a PARTNAME.
+warden_populate_by_name() {
+    mkdir -p /dev/block/by-name
+    for uev in /sys/class/block/vda*/uevent; do
+        [ -f "$uev" ] || continue
+        partname=""
+        devname=""
+        while IFS='=' read -r k v; do
+            case "$k" in
+                PARTNAME) partname="$v" ;;
+                DEVNAME)  devname="$v" ;;
+            esac
+        done < "$uev"
+        [ -n "$partname" ] && [ -n "$devname" ] \
+            && ln -sf "/dev/$devname" "/dev/block/by-name/$partname"
+    done
+}
+
+# Parse warden.slot= from the cmdline (whole-token, never substring) and
+# VALIDATE it — echoes "_a" or "_b", falling back to _a with a warning.
+warden_slot() {
+    slot="_a"
+    for tok in $(cat /proc/cmdline); do
+        case "$tok" in
+            warden.slot=*) slot="${tok#warden.slot=}" ;;
+        esac
+    done
+    case "$slot" in
+        _a|_b) ;;
+        *) echo "warden-lib: bad warden.slot='$slot', falling back to _a" >&2; slot="_a" ;;
+    esac
+    echo "$slot"
+}
diff --git a/qemu/rootfs/sbin/init b/qemu/rootfs/sbin/init
index 8eb2abd..c13b15c 100755
--- a/qemu/rootfs/sbin/init
+++ b/qemu/rootfs/sbin/init
@@ -13,33 +13,29 @@ mount -t proc proc /proc
 mount -t sysfs sysfs /sys
 mount -t tmpfs tmpfs /tmp
 
-# Fresh devtmpfs — repopulate the by-name contract (slotctl.rs depends on it).
-mkdir -p /dev/block/by-name
-for uev in /sys/class/block/vda*/uevent; do
-    [ -f "$uev" ] || continue
-    partname=""
-    devname=""
-    while IFS='=' read -r k v; do
-        case "$k" in
-            PARTNAME) partname="$v" ;;
-            DEVNAME)  devname="$v" ;;
-        esac
-    done < "$uev"
-    [ -n "$partname" ] && [ -n "$devname" ] \
-        && ln -sf "/dev/$devname" "/dev/block/by-name/$partname"
-done
+. /etc/warden-lib.sh
 
-# Slot (whole-token parse, same rule as stage 1).
-slot="_a"
-for tok in $(cat /proc/cmdline); do
-    case "$tok" in
-        warden.slot=*) slot="${tok#warden.slot=}" ;;
-    esac
-done
+# Fresh devtmpfs — repopulate the by-name contract; same VALIDATED slot rule
+# as stage 1 (shared helper, so the two can never drift).
+warden_populate_by_name
+slot="$(warden_slot)"
 
 # The device's matched mounts: persistent state and the slot's oem partition.
-mount -t ext4 /dev/block/by-name/userdata /userdata || echo "init: userdata mount failed"
-mount -t ext4 "/dev/block/by-name/oem${slot}" /oem || echo "init: oem${slot} mount failed"
+# Fail-fast: a scenario against an image whose userdata cannot mount would
+# otherwise burn its whole deadline before failing generically. warden.shell
+# still gets a shell for post-mortem.
+mount_fatal() {
+    if ! mount -t ext4 "/dev/block/by-name/$1" "$2"; then
+        echo "WARDEN-QEMU-MOUNT-FAILED $1"
+        if grep -qw warden.shell /proc/cmdline; then
+            echo "warden.shell: post-mortem shell (exit powers off)"
+            setsid cttyhack sh
+        fi
+        poweroff -f
+    fi
+}
+mount_fatal userdata /userdata
+mount_fatal "oem${slot}" /oem
 mkdir -p /userdata/warden
 
 # RS485: warden-modbus hardcodes /dev/ttyS4 at compile time; alias it to the
@@ -47,10 +43,13 @@ mkdir -p /userdata/warden
 [ -c /dev/ttyS0 ] && ln -sf /dev/ttyS0 /dev/ttyS4
 
 # Network: slirp user-mode net on eth0 (DHCP, fallback to QEMU's static map).
+# The fallback keys off the interface actually having an address — udhcpc
+# exiting 0 only proves a lease, not that the hook script applied it.
 ip link set lo up
 if [ -e /sys/class/net/eth0 ]; then
     ip link set eth0 up
-    if ! udhcpc -i eth0 -n -q -t 5 -T 2 >/dev/null 2>&1; then
+    udhcpc -i eth0 -n -q -t 5 -T 2 >/dev/null 2>&1 || true
+    if ! ip -4 addr show dev eth0 | grep -q 'inet '; then
         ip addr add 10.0.2.15/24 dev eth0 2>/dev/null
         ip route replace default via 10.0.2.2 dev eth0
         echo "nameserver 10.0.2.3" > /etc/resolv.conf
diff --git a/qemu/rs485-bridge/src/lib.rs b/qemu/rs485-bridge/src/lib.rs
index 92023ad..bb2fae2 100644
--- a/qemu/rs485-bridge/src/lib.rs
+++ b/qemu/rs485-bridge/src/lib.rs
@@ -24,6 +24,11 @@ use warden_sim::ModbusSlave;
 /// response timeout is orders of magnitude larger.
 pub const DEFAULT_GAP: Duration = Duration::from_millis(10);
 
+/// Accumulation cap: a Modbus RTU ADU is at most 256 bytes, so anything past
+/// 2x that without an inter-frame gap is a misbehaving master streaming
+/// continuously — drop the buffer instead of growing without bound.
+const MAX_PENDING: usize = 512;
+
 /// The shared bus: the slave plus its declared dimensions. The sim's register
 /// setters panic on out-of-range indices (deliberate test-harness semantics);
 /// the control channel must bounds-check first so a typo in a scenario script
@@ -64,7 +69,17 @@ pub fn pump_serial(
                 }
                 return Ok(());
             }
-            Ok(n) => buf.extend_from_slice(&chunk[..n]),
+            Ok(n) => {
+                buf.extend_from_slice(&chunk[..n]);
+                if buf.len() > MAX_PENDING {
+                    eprintln!(
+                        "rs485: {} bytes buffered with no inter-frame gap — discarding \
+                         (misbehaving master streaming continuously?)",
+                        buf.len()
+                    );
+                    buf.clear();
+                }
+            }
             Err(e)
                 if e.kind() == std::io::ErrorKind::WouldBlock
                     || e.kind() == std::io::ErrorKind::TimedOut =>
@@ -127,10 +142,10 @@ pub fn handle_control_line(line: &str, bus: &Bus) -> String {
     if words.next().is_some() {
         return format!("err trailing arguments after '{cmd}'");
     }
-    let bound = |cmd: &str| match cmd {
-        "holding" | "input" | "get-holding" => bus.regs,
-        _ => bus.bits,
-    };
+    // Each arm states its own bound (bus.regs for register space, bus.bits for
+    // bit space) INLINE — a previous string-keyed lookup defaulted silently to
+    // the bit bound, which would have handed a future `get-input` command the
+    // wrong range and reintroduced the out-of-range panic this check prevents.
     let mut s = bus.slave.lock().unwrap();
     match (cmd, arg) {
         ("ping", None) => "ok".into(),
@@ -152,39 +167,63 @@ pub fn handle_control_line(line: &str, bus: &Bus) -> String {
             }
             _ => format!("err bad exception code '{c}'"),
         },
-        ("holding" | "input" | "coil" | "discrete", Some(kv)) => {
-            let (addr, val) = match kv.split_once('=') {
-                Some((a, v)) => (parse_u16(a), parse_u16(v)),
-                None => (None, None),
-            };
-            match (addr, val) {
-                (Some(a), _) if (a as usize) >= bound(cmd) => {
-                    format!("err address {a} out of range (0..{})", bound(cmd))
+        ("holding" | "input", Some(kv)) => match parse_addr_val(kv, bus.regs) {
+            Ok((a, v)) => {
+                if cmd == "holding" {
+                    s.set_holding(a, v);
+                } else {
+                    s.set_input(a, v);
                 }
-                (Some(a), Some(v)) => {
-                    match cmd {
-                        "holding" => s.set_holding(a as usize, v),
-                        "input" => s.set_input(a as usize, v),
-                        "coil" => s.set_coil(a as usize, v != 0),
-                        _ => s.set_discrete(a as usize, v != 0),
-                    }
-                    "ok".into()
+                "ok".into()
+            }
+            Err(e) => e,
+        },
+        ("coil" | "discrete", Some(kv)) => match parse_addr_val(kv, bus.bits) {
+            Ok((a, v)) => {
+                if cmd == "coil" {
+                    s.set_coil(a, v != 0);
+                } else {
+                    s.set_discrete(a, v != 0);
                 }
-                _ => format!("err expected <addr>=<value>, got '{kv}'"),
+                "ok".into()
             }
-        }
-        ("get-holding" | "get-coil", Some(a)) => match parse_u16(a) {
-            Some(a) if (a as usize) >= bound(cmd) => {
-                format!("err address {a} out of range (0..{})", bound(cmd))
-            }
-            Some(a) if cmd == "get-holding" => format!("ok {}", s.holding(a as usize)),
-            Some(a) => format!("ok {}", u8::from(s.coil(a as usize))),
-            None => format!("err bad address '{a}'"),
+            Err(e) => e,
+        },
+        ("get-holding", Some(a)) => match parse_addr(a, bus.regs) {
+            Ok(a) => format!("ok {}", s.holding(a)),
+            Err(e) => e,
+        },
+        ("get-coil", Some(a)) => match parse_addr(a, bus.bits) {
+            Ok(a) => format!("ok {}", u8::from(s.coil(a))),
+            Err(e) => e,
         },
         _ => format!("err unknown or malformed command '{line}'"),
     }
 }
 
+/// Parse "<addr>=<value>" with the address bounds-checked against `bound`.
+fn parse_addr_val(kv: &str, bound: usize) -> Result<(usize, u16), String> {
+    let Some((a, v)) = kv.split_once('=') else {
+        return Err(format!("err expected <addr>=<value>, got '{kv}'"));
+    };
+    let (Some(a), Some(v)) = (parse_u16(a), parse_u16(v)) else {
+        return Err(format!("err expected <addr>=<value>, got '{kv}'"));
+    };
+    if (a as usize) >= bound {
+        return Err(format!("err address {a} out of range (0..{bound})"));
+    }
+    Ok((a as usize, v))
+}
+
+/// Parse a bare address, bounds-checked against `bound`.
+fn parse_addr(a: &str, bound: usize) -> Result<usize, String> {
+    match parse_u16(a) {
+        Some(v) if (v as usize) < bound => Ok(v as usize),
+        Some(v) => Err(format!("err address {v} out of range (0..{bound})")),
+        None => Err(format!("err bad address '{a}'")),
+    }
+}
+
 fn parse_u16(s: &str) -> Option<u16> {
     if let Some(h) = s.strip_prefix("0x").or_else(|| s.strip_prefix("0X")) {
         u16::from_str_radix(h, 16).ok()
@@ -200,10 +239,13 @@ mod tests {
     use std::time::Duration;
     use warden_sim::modbus::{crc_ok, read_holding};
 
-    // Test gap is larger than DEFAULT_GAP so a loaded CI runner cannot split
-    // a frame that the test wrote in two deliberate chunks.
-    const GAP: Duration = Duration::from_millis(25);
-    const SETTLE: Duration = Duration::from_millis(100);
+    // Test gap is much larger than DEFAULT_GAP so a loaded CI runner cannot
+    // split a frame the test wrote in two deliberate chunks: the 2ms
+    // inter-chunk pause has a 60x margin against the 120ms dispatch gap
+    // (25ms gave only 12.5x and was flagged as a flake risk on contended
+    // 2-vCPU hosted runners).
+    const GAP: Duration = Duration::from_millis(120);
+    const SETTLE: Duration = Duration::from_millis(400);
 
     fn bus() -> Bus {
         let b = Bus::new(1, 16, 16);
@@ -221,7 +263,9 @@ mod tests {
     }
 
     fn read_reply(master: &UnixStream) -> Vec<u8> {
-        master.set_read_timeout(Some(Duration::from_secs(2))).unwrap();
+        master
+            .set_read_timeout(Some(Duration::from_secs(2)))
+            .unwrap();
         let mut buf = [0u8; 256];
         let n = (&*master).read(&mut buf).expect("expected a reply frame");
         buf[..n].to_vec()
@@ -274,7 +318,10 @@ mod tests {
         with_pump(&s, |master| {
             (&*master).write_all(&read_holding(1, 2, 1)).unwrap();
             master
-                .set_read_timeout(Some(Duration::from_millis(200)))
+                // Well past GAP: the dropped frame must have been dispatched
+                // (and answered with silence) before the next request is
+                // written, or the two would merge in the pending buffer.
+                .set_read_timeout(Some(Duration::from_millis(500)))
                 .unwrap();
             let mut buf = [0u8; 16];
             assert!(
diff --git a/qemu/rs485-bridge/src/main.rs b/qemu/rs485-bridge/src/main.rs
index f9ac955..6e1cc7e 100644
--- a/qemu/rs485-bridge/src/main.rs
+++ b/qemu/rs485-bridge/src/main.rs
@@ -2,10 +2,14 @@
 //! there); this file only parses arguments, connects sockets, and spawns the
 //! control listener.
 //!
-//! Typical use (matches qemu/run.sh --rs485):
+//! Typical use (matches qemu/run.sh --rs485). Put the sockets in a private
+//! per-run directory (mktemp -d) — short (AF_UNIX caps paths at ~108 chars)
+//! and not guessable/pre-creatable by other local users, unlike a fixed
+//! /tmp name:
 //!
-//!   qemu/run.sh --kernel ... --rs485 /tmp/warden-rs485.sock &
-//!   rs485-bridge --serial /tmp/warden-rs485.sock --control /tmp/warden-rs485-ctl.sock
+//!   d=$(mktemp -d /tmp/rs485.XXXXXX)
+//!   qemu/run.sh --kernel ... --rs485 "$d/serial.sock" &
+//!   rs485-bridge --serial "$d/serial.sock" --control "$d/ctl.sock"
 
 use std::io::{BufRead, BufReader, Write};
 use std::os::unix::net::{UnixListener, UnixStream};
@@ -31,10 +35,19 @@ fn main() {
 
     let mut args = std::env::args().skip(1);
     while let Some(a) = args.next() {
-        let mut val = |name: &str| args.next().unwrap_or_else(|| {
-            eprintln!("{name} needs a value");
-            usage()
-        });
+        let mut val = |name: &str| {
+            let v = args.next().unwrap_or_else(|| {
+                eprintln!("{name} needs a value");
+                usage()
+            });
+            // A following flag means the value was omitted — report the real
+            // problem instead of swallowing the flag as a bogus value.
+            if v.starts_with("--") {
+                eprintln!("{name} needs a value, got flag '{v}'");
+                usage()
+            }
+            v
+        };
         match a.as_str() {
             "--serial" => serial = Some(val("--serial")),
             "--control" => control = Some(val("--control")),
@@ -55,14 +68,32 @@ fn main() {
     let bus: &'static Bus = Box::leak(Box::new(Bus::new(address, regs, bits)));
 
     if let Some(path) = control {
-        let _ = std::fs::remove_file(&path); // stale socket from a previous run
+        // Clear a stale socket from a previous run. A failure here that is not
+        // "nothing to remove" (e.g. someone else's file behind /tmp's sticky
+        // bit) will make the bind below fail — surface both errors.
+        let removed = std::fs::remove_file(&path);
         let listener = UnixListener::bind(&path).unwrap_or_else(|e| {
             eprintln!("FATAL: cannot bind control socket {path}: {e}");
+            if let Err(re) = removed {
+                if re.kind() != std::io::ErrorKind::NotFound {
+                    eprintln!("       (removing the pre-existing file also failed: {re})");
+                }
+            }
             exit(1);
         });
         eprintln!("rs485: control socket at {path}");
         std::thread::spawn(move || {
-            for conn in listener.incoming().flatten() {
+            // Explicit error handling: `.flatten()` would turn a persistent
+            // accept() failure (fd exhaustion etc.) into a silent hot loop.
+            for conn in listener.incoming() {
+                let conn = match conn {
+                    Ok(c) => c,
+                    Err(e) => {
+                        eprintln!("rs485: control accept failed: {e} — backing off");
+                        std::thread::sleep(Duration::from_millis(200));
+                        continue;
+                    }
+                };
                 let reader = BufReader::new(conn.try_clone().expect("clone control conn"));
                 let mut writer = conn;
                 for line in reader.lines() {
diff --git a/qemu/run.sh b/qemu/run.sh
index 7d895c5..a40dde5 100755
--- a/qemu/run.sh
+++ b/qemu/run.sh
@@ -18,9 +18,9 @@
 #   --qmp SOCK         QMP unix socket (screendump, input-send-event, quit)
 #   --display MODE     off (default, -nographic) | on (gtk window) | headless
 #                      (virtio-gpu without a window; screendump via --qmp)
-#   --ssh-port N       hostfwd 127.0.0.1:N -> guest :22   (default 2222)
-#   --http-port N      hostfwd 127.0.0.1:N -> guest :80   (default 8080)
-#   --api-port N       hostfwd 127.0.0.1:N -> guest :28443 (default 28443)
+#   --ssh-port N       hostfwd 127.0.0.1:N -> guest :22   (default 2222; 0 disables)
+#   --http-port N      hostfwd 127.0.0.1:N -> guest :80   (default 8080; 0 disables)
+#   --api-port N       hostfwd 127.0.0.1:N -> guest :28443 (default 28443; 0 disables)
 #   --shell            interactive shell in the guest instead of daemon hold
 set -euo pipefail
 
@@ -76,13 +76,19 @@ fi
 # NOTE: never add `earlyprintk` — the config's DEBUG_UART_PHYS is the RV1106's
 # 0xff4c0000, which does not exist on -M virt.
 APPEND="console=ttyAMA0 rdinit=/init"
+# Port 0 disables a forward — a boot smoke needs no host ports and must not
+# fail on a busy default port.
+NETDEV="user,id=n0"
+[ "$SSH_PORT" != 0 ]  && NETDEV="$NETDEV,hostfwd=tcp:127.0.0.1:${SSH_PORT}-:22"
+[ "$HTTP_PORT" != 0 ] && NETDEV="$NETDEV,hostfwd=tcp:127.0.0.1:${HTTP_PORT}-:80"
+[ "$API_PORT" != 0 ]  && NETDEV="$NETDEV,hostfwd=tcp:127.0.0.1:${API_PORT}-:28443"
 ARGS=(
   -M "virt,highmem=off" -cpu cortex-a7 -smp 1 -m 256M
   # virtio-mmio defaults to the legacy (0.9) transport; virtio-gpu and
   # virtio-input are VERSION_1-only devices and never bind without this.
   -global "virtio-mmio.force-legacy=false"
   -kernel "$KERNEL" -initrd "$INITRD"
-  -netdev "user,id=n0,hostfwd=tcp:127.0.0.1:${SSH_PORT}-:22,hostfwd=tcp:127.0.0.1:${HTTP_PORT}-:80,hostfwd=tcp:127.0.0.1:${API_PORT}-:28443"
+  -netdev "$NETDEV"
   -device "virtio-net-device,netdev=n0"
   -no-reboot
 )
diff --git a/qemu/tests/boot-smoke.sh b/qemu/tests/boot-smoke.sh
index c8ab969..15338e4 100755
--- a/qemu/tests/boot-smoke.sh
+++ b/qemu/tests/boot-smoke.sh
@@ -27,13 +27,15 @@ command -v qemu-system-arm >/dev/null || {
 LOG="$(mktemp "${TMPDIR:-/tmp}/warden-qemu-smoke.XXXXXX")"
 trap 'rm -f "$LOG"' EXIT
 
-# NOTE: never pass `earlyprintk` — the config's DEBUG_UART_PHYS is the RV1106's
-# 0xff4c0000, which does not exist on -M virt.
-timeout 180 qemu-system-arm \
-  -M virt,highmem=off -cpu cortex-a7 -smp 1 -m 256M \
-  -kernel "$ZIMAGE" -initrd "$INITRD" \
-  -append "console=ttyAMA0 rdinit=/init" \
-  -nographic -no-reboot </dev/null | tee "$LOG" || {
+# Delegate the qemu invocation to run.sh (--no-disk) so the machine shape
+# (-M virt,highmem=off, cpu, memory, virtio topology) lives in exactly one
+# place — the two hand-copied invocations had already drifted once.
+# timeout -k: a wedged qemu that ignores SIGTERM gets SIGKILLed 10s later
+# instead of holding the job until the workflow-level timeout.
+timeout -k 10 180 bash "$QDIR/run.sh" \
+  --kernel "$ZIMAGE" --initrd "$INITRD" --no-disk \
+  --ssh-port 0 --http-port 0 --api-port 0 \
+  </dev/null | tee "$LOG" || {
     echo "FATAL: qemu exited non-zero (or hung until the 180s timeout)" >&2
     exit 1
   }
diff --git a/qemu/tests/portal-scenario.sh b/qemu/tests/portal-scenario.sh
index bd185be..89e101f 100755
--- a/qemu/tests/portal-scenario.sh
+++ b/qemu/tests/portal-scenario.sh
@@ -62,11 +62,19 @@ python3 "$FLARE_EDGE/tools/mock-flare-portal.py" \
   --port "$PORT" --device "$DEVICE_ID:$API_KEY" \
   --wfw "$WORK/offer.wfw" > "$WORK/mock.log" 2>&1 &
 MOCK_PID=$!
+mock_ready=0
 for _ in $(seq 1 50); do
-  curl -so /dev/null "http://127.0.0.1:$PORT/" && break
+  curl -so /dev/null "http://127.0.0.1:$PORT/" && { mock_ready=1; break; }
   kill -0 "$MOCK_PID" 2>/dev/null || { echo "FATAL: mock portal died:" >&2; cat "$WORK/mock.log" >&2; exit 1; }
   sleep 0.2
 done
+# The loop must not fall through silently: an alive-but-unresponsive mock
+# would otherwise surface 420s later as an unrelated assertion timeout.
+[ "$mock_ready" = 1 ] || {
+  echo "FATAL: mock portal never answered on :$PORT within 10s" >&2
+  tail -10 "$WORK/mock.log" >&2
+  exit 1
+}
 echo "== mock portal on :$PORT, device $DEVICE_ID"
 
 # 2. image seeded with the portal URL + credentials.
@@ -79,11 +87,34 @@ bash "$QDIR/mkimage.sh" \
   --state "flare.api_key=$API_KEY" \
   --state "flare.site=qemu-devsim"
 
-# 3. boot the VM headless (daemons run; console log to file).
-bash "$QDIR/run.sh" --kernel "$ZIMAGE" \
-  --ssh-port $((PORT + 1)) --http-port $((PORT + 2)) --api-port $((PORT + 3)) \
-  > "$WORK/console.log" 2>&1 &
-QEMU_PID=$!
+# 3. boot the VM headless (daemons run; console log to file). Random hostfwd
+#    ports can collide with another process — detect the early qemu bind
+#    failure and retry with a fresh base rather than failing spuriously.
+QEMU_PID=""
+for _attempt in 1 2 3; do
+  VMBASE=$((20000 + RANDOM % 20000))
+  : > "$WORK/console.log"
+  bash "$QDIR/run.sh" --kernel "$ZIMAGE" \
+    --ssh-port "$VMBASE" --http-port $((VMBASE + 1)) --api-port $((VMBASE + 2)) \
+    > "$WORK/console.log" 2>&1 &
+  QEMU_PID=$!
+  sleep 3
+  if kill -0 "$QEMU_PID" 2>/dev/null; then
+    break
+  fi
+  if grep -aq 'Could not set up host forwarding' "$WORK/console.log"; then
+    echo "== hostfwd port collision on base $VMBASE — retrying"
+    QEMU_PID=""
+    continue
+  fi
+  echo "FATAL: VM died at launch:" >&2
+  tail -20 "$WORK/console.log" >&2
+  exit 1
+done
+if [ -z "$QEMU_PID" ] || ! kill -0 "$QEMU_PID" 2>/dev/null; then
+  echo "FATAL: could not launch the VM after 3 port attempts" >&2
+  exit 1
+fi
 
 # 4. assert: rootfs up, and the portal saw — from OUR device id — an
 #    authenticated check-in, the firmware desired-state pull, and the signed
@@ -99,6 +130,11 @@ while [ $SECONDS -lt $deadline ]; do
   grep -aq "GET /api/v1/devices/$DEVICE_ID/firmware -> 200" "$WORK/mock.log" && ok_fw=1
   grep -aq "GET /api/v1/devices/$DEVICE_ID/firmware/assets/.* -> 200" "$WORK/mock.log" && ok_asset=1
   [ $ok_report -eq 1 ] && [ $ok_fw -eq 1 ] && [ $ok_asset -eq 1 ] && break
+  grep -aq 'WARDEN-QEMU-MOUNT-FAILED' "$WORK/console.log" && {
+    echo "FATAL: guest partition mount failed (bad image?):" >&2
+    grep -a 'WARDEN-QEMU-MOUNT-FAILED' "$WORK/console.log" >&2
+    exit 1
+  }
   kill -0 "$QEMU_PID" 2>/dev/null || { echo "FATAL: VM exited early" >&2; tail -30 "$WORK/console.log" >&2; exit 1; }
   sleep 2
 done
diff --git a/qemu/tests/qmp.py b/qemu/tests/qmp.py
index 5c4668d..0737cfb 100755
--- a/qemu/tests/qmp.py
+++ b/qemu/tests/qmp.py
@@ -11,7 +11,7 @@ import sys
 import time
 
 
-def rpc(sock, obj):
+def rpc(sock, sock_file, obj):
     sock.sendall((json.dumps(obj) + "\n").encode())
     while True:
         line = sock_file.readline()
@@ -29,15 +29,20 @@ def main():
     if len(sys.argv) < 3:
         sys.exit(__doc__)
     path, cmd = sys.argv[1], sys.argv[2]
-    global sock_file
+    need = {"screendump": 4, "tap": 5, "quit": 3}
+    if cmd not in need:
+        sys.exit(f"unknown command {cmd}\n{__doc__}")
+    if len(sys.argv) < need[cmd]:
+        sys.exit(f"{cmd}: missing argument(s)\n{__doc__}")
+
     s = socket.socket(socket.AF_UNIX)
     s.connect(path)
-    sock_file = s.makefile("r")
-    sock_file.readline()  # greeting banner
-    rpc(s, {"execute": "qmp_capabilities"})
+    f = s.makefile("r")
+    f.readline()  # greeting banner
+    rpc(s, f, {"execute": "qmp_capabilities"})
 
     if cmd == "screendump":
-        rpc(s, {"execute": "screendump", "arguments": {"filename": sys.argv[3]}})
+        rpc(s, f, {"execute": "screendump", "arguments": {"filename": sys.argv[3]}})
     elif cmd == "tap":
         x, y = int(sys.argv[3]), int(sys.argv[4])
         press = [
@@ -46,16 +51,14 @@ def main():
             {"type": "btn", "data": {"down": True, "button": "left"}},
         ]
         release = [{"type": "btn", "data": {"down": False, "button": "left"}}]
-        rpc(s, {"execute": "input-send-event", "arguments": {"events": press}})
+        rpc(s, f, {"execute": "input-send-event", "arguments": {"events": press}})
         # Hold the press across several LVGL indev poll periods (33 ms each):
         # an instantaneous press+release lands inside one poll and no click
         # is ever registered.
         time.sleep(0.2)
-        rpc(s, {"execute": "input-send-event", "arguments": {"events": release}})
+        rpc(s, f, {"execute": "input-send-event", "arguments": {"events": release}})
     elif cmd == "quit":
         s.sendall(b'{"execute":"quit"}\n')
-    else:
-        sys.exit(f"unknown command {cmd}")
 
 
 if __name__ == "__main__":
diff --git a/qemu/tests/ui-shot.sh b/qemu/tests/ui-shot.sh
index 82250cc..c58ee83 100755
--- a/qemu/tests/ui-shot.sh
+++ b/qemu/tests/ui-shot.sh
@@ -1,9 +1,10 @@
 #!/usr/bin/env bash
 # Display + touch scenario: boot the VM headless with virtio-gpu, wait for the
-# LVGL UI (fbdev build) to start, screendump over QMP, inject an absolute
-# touch tap (virtio-tablet), screendump again. Asserts the first frame is
-# non-blank; reports (does not assert) whether the tap changed pixels — the
-# device-level analogue of flare-edge's Xvfb/xdotool sim-test.sh.
+# LVGL UI (fbdev build) to render a real frame, then inject an absolute touch
+# tap on the Metrics tab (virtio-tablet) and ASSERT the frame changed — the
+# device-level analogue of flare-edge's Xvfb/xdotool sim-test.sh. Readiness is
+# polled from screendumps on bounded deadlines, never guessed with fixed
+# sleeps: TCG renders CPU-bound and a loaded host can be arbitrarily slow.
 #
 # FAILS CLOSED on missing prerequisites.
 #
@@ -40,11 +41,30 @@ trap cleanup EXIT
 bash "$QDIR/mkinitramfs.sh"
 bash "$QDIR/mkimage.sh"
 
-PORT=$((21000 + RANDOM % 20000))
-bash "$QDIR/run.sh" --kernel "$ZIMAGE" --display headless --qmp "$WORK/qmp.sock" \
-  --ssh-port "$PORT" --http-port $((PORT + 1)) --api-port $((PORT + 2)) \
-  > "$WORK/console.log" 2>&1 &
-QEMU_PID=$!
+# Random hostfwd ports can collide — detect qemu's early bind failure and
+# retry with a fresh base rather than failing spuriously.
+for _attempt in 1 2 3; do
+  PORT=$((21000 + RANDOM % 20000))
+  : > "$WORK/console.log"
+  bash "$QDIR/run.sh" --kernel "$ZIMAGE" --display headless --qmp "$WORK/qmp.sock" \
+    --ssh-port "$PORT" --http-port $((PORT + 1)) --api-port $((PORT + 2)) \
+    > "$WORK/console.log" 2>&1 &
+  QEMU_PID=$!
+  sleep 3
+  kill -0 "$QEMU_PID" 2>/dev/null && break
+  if grep -aq 'Could not set up host forwarding' "$WORK/console.log"; then
+    echo "== hostfwd port collision on base $PORT — retrying"
+    QEMU_PID=""
+    continue
+  fi
+  echo "FATAL: VM died at launch:" >&2
+  tail -20 "$WORK/console.log" >&2
+  exit 1
+done
+if [ -z "$QEMU_PID" ] || ! kill -0 "$QEMU_PID" 2>/dev/null; then
+  echo "FATAL: could not launch the VM after 3 port attempts" >&2
+  exit 1
+fi
 
 deadline=$((SECONDS + 120))
 while [ $SECONDS -lt $deadline ]; do
@@ -57,36 +77,55 @@ grep -aq 'init: starting warden-ui' "$WORK/console.log" || {
   tail -25 "$WORK/console.log" >&2
   exit 1
 }
-sleep 8   # let LVGL render the first frames
 
 qmp() { python3 "$HERE/qmp.py" "$WORK/qmp.sock" "$@"; }
 
-qmp screendump "$WORK/shot1.ppm"
+# Frame is "real" once it has more than a handful of distinct colors (a blank
+# or console-only frame has very few).
+frame_rendered() { # $1 = ppm path
+  python3 - "$1" <<'EOF'
+import sys
+data = open(sys.argv[1], "rb").read()
+parts = data.split(b"\n", 3)               # P6 header: magic, dims, maxval, raw RGB
+pixels = parts[3] if len(parts) == 4 else b""
+distinct = len(set(pixels[i:i+3] for i in range(0, min(len(pixels), 3*720*720), 3)))
+print(f"{sys.argv[1]}: {len(pixels)} bytes, {distinct} distinct colors")
+sys.exit(0 if distinct > 16 else 1)
+EOF
+}
+
+# Poll for the first rendered frame (bounded, no guessed sleep).
+rendered=0
+deadline=$((SECONDS + 90))
+while [ $SECONDS -lt $deadline ]; do
+  qmp screendump "$WORK/shot1.ppm"
+  if frame_rendered "$WORK/shot1.ppm"; then rendered=1; break; fi
+  sleep 3
+done
+[ "$rendered" = 1 ] || {
+  echo "FATAL: UI never rendered a non-blank frame within 90s" >&2
+  exit 1
+}
+
 # Tap the "Metrics" tab: pixel (373,40) of 720x720 scaled to the QMP absolute
-# range 0..32767 — switching tabs must repaint the content area.
+# range 0..32767 — switching tabs must repaint the content area. Poll for the
+# repaint rather than guessing a delay.
 qmp tap 16975 1820
-sleep 3
-qmp screendump "$WORK/shot2.ppm"
+changed=0
+deadline=$((SECONDS + 30))
+while [ $SECONDS -lt $deadline ]; do
+  sleep 2
+  qmp screendump "$WORK/shot2.ppm"
+  if ! cmp -s "$WORK/shot1.ppm" "$WORK/shot2.ppm"; then changed=1; break; fi
+done
 
 mkdir -p "$OUTDIR"
 cp "$WORK/shot1.ppm" "$OUTDIR/ui-shot1.ppm"
-cp "$WORK/shot2.ppm" "$OUTDIR/ui-shot2.ppm"
+cp "$WORK/shot2.ppm" "$OUTDIR/ui-shot2.ppm" 2>/dev/null || true
 
-# Non-blank: more than one distinct pixel value in the raw PPM payload.
-python3 - "$WORK/shot1.ppm" <<'EOF'
-import sys
-data = open(sys.argv[1], "rb").read()
-# P6 header: magic, dims, maxval, then raw RGB
-parts = data.split(b"\n", 3)
-pixels = parts[3] if len(parts) == 4 else b""
-distinct = len(set(pixels[i:i+3] for i in range(0, min(len(pixels), 3*720*720), 3)))
-print(f"shot1: {len(pixels)} bytes of pixels, {distinct} distinct colors")
-sys.exit(0 if distinct > 1 else 1)
-EOF
-
-if cmp -s "$WORK/shot1.ppm" "$WORK/shot2.ppm"; then
-  echo "FATAL: tapping the Metrics tab did not change the frame — touch is not reaching the UI" >&2
+[ "$changed" = 1 ] || {
+  echo "FATAL: tapping the Metrics tab did not change the frame within 30s — touch is not reaching the UI" >&2
   exit 1
-fi
+}
 echo "tap on the Metrics tab repainted the frame (touch reached the UI)"
 echo "UI-SHOT-PASS (screenshots in $OUTDIR/ui-shot{1,2}.ppm)"