fix(vendor-ie): explicit 255-byte overflow policy — truncate, report, never silently diverge

.github/workflows/build-package.yml

@@ -1025,6 +1025,19 @@ jobs:
           python3 -m pip install --break-system-packages playwright
           python3 -m playwright install --with-deps chromium
 
+      # Dependency-resolution smoke: apk3 (the only tooling that reads 25.12
+      # indexes) must be able to select every name the package declares from
+      # the standard 25.12.x feed set. This is the gate the September #552
+      # breakage class would have tripped: a feed rebuild changed how the
+      # iptables family is provided, and nothing between "artifact builds"
+      # and "a bench VM cannot install it" noticed. Runs before the
+      # happy-path suite because it is seconds, not minutes.
+      - name: Apk dependency-resolution smoke
+        run: |
+          set -euo pipefail
+          export APK=$(ls /var/tmp/hp-artifact/*.apk | head -1)
+          bash tests/packaging/apk-install-resolution_test.sh
+
       # --strict is deliberate, not decoration: this artifact is built from a tip
       # that carries upstream #541 (/session-state) and a portal bundle that ships
       # the in-page renewal CTA (#60), so those surfaces are EXPECTED here and

.github/workflows/gitleaks.yml

@@ -0,0 +1,12 @@
+name: gitleaks
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+  schedule:
+    - cron: "13 6 * * *"
+
+jobs:
+  gitleaks:
+    uses: Amperstrand/.github/.github/workflows/gitleaks.yml@main

.github/workflows/test.yml

@@ -117,6 +117,32 @@ jobs:
           # negative control fails if the comparison ever goes back to equality.
           bash tests/uci-defaults-setup-marker-order_test.sh
 
+      - name: Exactly one owner of :8090, and LuCI alone on :8080
+        run: |
+          # Offline (no router, no SDK, no network): the portal-staged board
+          # (uhttpd.admin, written by the feed's 92-tollgate-admin-setup) is the
+          # ONE section that may claim :8090, and the module writes no :8090
+          # listener of its own. It used to: a brand-gated legacy configUI
+          # writer created a second section on the port, so two admin UIs bound
+          # the same port and one of them disappeared. The suite drives the
+          # shipped uci-defaults script end to end through the four install
+          # scenarios - fresh, upgrade, reinstall-over-marker, and a box whose
+          # legacy section had its listeners stripped by an older build (which
+          # is why the stale section must be DELETED, not port-stripped) - and
+          # carries detector controls so a passing assertion cannot be vacuous.
+          bash tests/packaging/configui-8090-single-owner_test.sh
+
+      - name: No re-brand literal anywhere in the tree
+        run: |
+          # Offline and tree-only: upstream carries no commercial re-brand's
+          # name - not in the setup script, not in the docs, not in the
+          # CHANGELOG. The gutter fails the moment a literal reappears, and
+          # scans itself: the pattern is a bracket expression, so this file does
+          # not carry the name it bans. A planted-occurrence control proves the
+          # scan is not inert. This is a working-tree gutter, not a history
+          # rewrite: the literal legitimately survives in old commits.
+          bash tests/packaging/rebrand-literal-gutter_test.sh
+
       - name: Admin TLS identity (HTTPS is not inherited from the image)
         run: |
           # Offline (no router, no SDK, no network): the install path must
@@ -163,13 +189,29 @@ jobs:
           # bridge's port list cleared so a port is never on two layer-2
           # domains. The same-version verify/repair path re-asserts the
           # placement (a factory reset is repaired), the module keep-list
-          # still carries /etc/config/network, and all four guard fragments
-          # are pinned byte-identical to main and still br-lan-scoped - that
-          # literal is what keeps a guest off :8090/:8443 and LuCI while a
-          # br-private client gets in. A neutralised-writer negative control
-          # must go red.
+          # still carries /etc/config/network, and the guard fragments are
+          # pinned by digest (three byte-identical to main; the :2121
+          # backend-firewall guard updated on purpose so the owner network can
+          # read the admin board's data) and still br-lan-scoped - that literal
+          # is what keeps a guest off :8090/:8443 and LuCI while a br-private
+          # client gets in. A neutralised-writer negative control must go red.
           bash tests/uci-defaults-lan-private-wired_test.sh
 
+      - name: The backend API answers the owner network (admin board data path)
+        run: |
+          # Offline (no router, no daemon, no network): the :2121 API must be
+          # reachable from br-private as well as br-lan. The admin board is
+          # deliberately kept OFF the captive bridge, so br-private is the one
+          # network it is administered from - and the board is a thin shell
+          # whose every panel reads pricing/whoami/balance/ln-invoice from
+          # :2121. With the exemption missing, the board rendered on a freshly
+          # flashed GL-MT3000 (2026-10-04, alpha4-pre21) while each data call
+          # died with "TypeError: NetworkError when attempting to fetch
+          # resource" and retried forever. The suite pins the exemption set on
+          # both protocol families, that the rule is still a drop, that no
+          # foreign interface is exempted, and that the fragment compiles.
+          bash tests/packaging/backend-api-owner-network_test.sh
+
       - name: One device code (hostname, captive SSID and private SSID agree)
         run: |
           # Offline (no router, no SDK, no network): the router's identity is

.gitignore

@@ -85,3 +85,9 @@ scripts/token-recovery/token-recovery
 
 # pytest bytecode from local (possibly root-owned) cloud-lab runs
 tests/cloud-lab/__pycache__/
+scripts/__pycache__/
+tests/cloud-lab/conformance/__pycache__/
+
+# Conformance lane runtime state (tokens stashed between phases, generated
+# config, verdict parts, evidence logs) — see tests/cloud-lab/conformance/.
+tests/cloud-lab/.conformance/

.ngit/README.md

@@ -148,6 +148,19 @@ DQ05, `embedded-act` runner, `ghcr.io/catthehacker/ubuntu:act-latest`) with
   falls back to `0` when history is unavailable; tagged releases are unaffected
   because their version is the tag name. `GOFLAGS=-buildvcs=false` for the same
   reason.
+  This also removes the **commit timestamp**, which the reproducible-build pin
+  (#383) needs for `SOURCE_DATE_EPOCH`. `scripts/ngit-commit-epoch.sh` is the one
+  derivation: the commit time from local history where there is one, and
+  otherwise a depth-1 fetch of *exactly that commit* from the ngit mirror the
+  release is built from — so the value is a property of the commit, not of the
+  run. Stage 1 (`determine-versioning`, `build-portal`) and the shards'
+  `resolve-inputs` all call it. When neither source answers, the job **fails**
+  instead of substituting the runner clock: the earlier cascade had a
+  `date +%s` last resort, and because the coordinator's synthesized push payload
+  carries `head_commit` with no timestamp it took that branch on *every* run
+  (`source: job clock (NOT commit-derived …)` at `ca5d07a2`), so two builds of
+  one commit stamped different `BuildTime` strings and the published bytes could
+  never be reproduced or compared.
 * **Blossom reachability from the runner.** `blossom.primal.net`,
   `blossom.psbt.me`, `blossom2.orangesync.tech` and `drive.cashu.email` answer;
   `blossom1.orangesync.tech` times out (25 s). The server list is left as the

.ngit/act/workflows/build-package-binaries.yml

@@ -110,31 +110,38 @@ jobs:
             echo "release_channel=dev" >> "$GITHUB_OUTPUT"
           fi
 
-      - name: Derive SOURCE_DATE_EPOCH from the source commit
+      - name: Resolve SOURCE_DATE_EPOCH from the source commit
         # The reproducible-build pin from #383: every embedded timestamp —
         # BuildTime in the binaries, ipk archive mtimes, portal output —
         # derives from the source commit, never from the wall clock, so a
-        # rebuild of one commit yields the same bytes. The act workspace has
-        # no usable .git, so the derivation is env-first like repro-check.yml:
-        # triggering commit's timestamp, job-start clock as the last resort
-        # (which is then recorded in the stage-1 rendezvous record below, so
-        # stage 2 still packages with exactly the epoch the binaries used).
+        # rebuild of one commit yields the same bytes.
+        #
+        # scripts/ngit-commit-epoch.sh is the single derivation, shared with the
+        # build-portal job below and with the shards' resolve-inputs: it reads
+        # the commit time from the local history when there is one, and
+        # otherwise fetches exactly this commit (depth 1) from the ngit mirror
+        # the release is built from — an act job checkout has no git metadata,
+        # `git log` fails there, and the coordinator's synthesized push payload
+        # carries `head_commit` without a timestamp, so the previous cascade
+        # landed on `date +%s` on EVERY ngit run and stamped the binaries with
+        # the runner's clock (observed at ca5d07a2: "source: job clock (NOT
+        # commit-derived ...)"). That is exactly the failure this lane exists to
+        # prevent: two builds of one commit produced different BuildTime strings,
+        # so the published bytes could never be reproduced or compared.
+        #
+        # No wall-clock fallback on purpose. When neither source is available
+        # the script exits non-zero and this step fails with it: one clearly
+        # red run is better than a release whose artifacts silently cannot be
+        # rebuilt identically.
         id: source-epoch
         shell: bash
         run: |
+          set -euo pipefail
           : ${GITHUB_OUTPUT:=/tmp/github_output}
-          EPOCH="$(git -C "$GITHUB_WORKSPACE" log -1 --format=%ct HEAD 2>/dev/null || true)"
-          SOURCE="git"
-          if [ -z "$EPOCH" ] && [ -n "${{ github.event.head_commit.timestamp }}" ]; then
-            EPOCH="$(date -u -d "${{ github.event.head_commit.timestamp }}" +%s)"
-            SOURCE="head_commit.timestamp"
-          fi
-          if [ -z "$EPOCH" ]; then
-            EPOCH="$(date +%s)"
-            SOURCE="job clock (NOT commit-derived — binaries will not rebuild identically)"
-          fi
+          EPOCH=$(bash scripts/ngit-commit-epoch.sh "$GITHUB_SHA")
+          case "$EPOCH" in ''|*[!0-9]*) echo "ERROR: not an epoch: '$EPOCH'" >&2; exit 1 ;; esac
           echo "source_date_epoch=$EPOCH" >> "$GITHUB_OUTPUT"
-          echo "SOURCE_DATE_EPOCH=$EPOCH (source: $SOURCE)"
+          echo "SOURCE_DATE_EPOCH=$EPOCH ($(date -u -d "@$EPOCH" '+%Y-%m-%d %H:%M:%S UTC'))"
 
       - name: Report
         run: |
@@ -303,8 +310,9 @@ jobs:
           set -euo pipefail
           : ${GITHUB_OUTPUT:=/tmp/github_output}
           # The epoch rides the record so stage 2 packages with exactly the
-          # epoch these binaries were stamped with — even when the last-resort
-          # clock source produced it.
+          # epoch these binaries were stamped with. It is commit-derived (see
+          # the source-epoch step), so this record is stable across re-runs at
+          # one commit rather than reporting whenever the run happened.
           RECORD_JSON=$(printf '%s' "$HASHES_JSON" | jq -c --argjson epoch "$SOURCE_DATE_EPOCH" '. + {epoch: $epoch}')
           out=$(nak event --sec "$NSEC_HEX" -k 30078 \
             --tag "d=tollgate-build/${BUILD_ID}/binaries" \

AGENTS.md

@@ -118,8 +118,10 @@ than inventing new harnesses.
 - **gonuts-tollgate is our fork to maintain.** Upstream `elnosh/gonuts`
   is dead (last release v0.4.2, 2025); we carry ~40 patches. Every
   wallet-level fix lands in `OpenTollGate/gonuts-tollgate` first, is
-  tagged, then bumped here via the `replace` directive (three `go.mod`
-  files). Never fix a wallet bug by patching around the fork locally.
+  tagged, then bumped here via the `replace` directive — the require lands
+  in every nested module that carries it: `src/`, `src/cli`, `src/merchant`
+  and `src/tollwallet`, four `go.mod` files today. Never fix a wallet bug
+  by patching around the fork locally.
 - **bbolt persistence.** Keyset records (which own derivation counters)
   are nested under mint-URL-named buckets; the DB has no transactions
   spanning "fetch keysets + swap + save proofs". This is why counter
@@ -141,6 +143,47 @@ than inventing new harnesses.
   breaks that (e.g. cdk-go FFI on MIPS) belongs behind the sidecar, not
   in-process.
 
+## Hardware and VM testing (labgrid)
+
+All router- and VM-based testing is coordinated through **labgrid**
+(coordinator `ai-legion:20408`). Do not drive lab hardware ad hoc: reserve
+through places (`labgrid-client -p <place> acquire` … `release`), and treat
+a place held by someone else as theirs. Note some hosts carry a stale
+`LG_COORDINATOR` pointing at a dead address — use the hostname form:
+
+```bash
+export LG_COORDINATOR=ai-legion:20408
+labgrid-client places          # inventory + comments say what each seat is
+labgrid-client who             # current holders
+```
+
+The lab's single source of truth is the private
+**`Amperstrand/conwrt-bench`** repo (ADR-0005): `registry/` for devices,
+`labgrid/` for place seeds and examples, `docs/decisions/` for the why,
+sops for secrets. `conwrt-lab` is retired — do not add data there. Rules
+that every hardware-touching change follows:
+
+1. Never hardcode device IPs, MACs or serial paths — resolve from the
+   registry or a labgrid place.
+2. Access hardware through labgrid places (`ssh|console|power`), with
+   acquire/release for exclusivity during a test.
+3. Flashing and adoption go through conwrt tooling
+   (`dut_recover.py --from-lab`, `bench_net.py`), which updates the
+   registry.
+4. A state change ends with a registry commit — flashed, moved or
+   adopted devices must be reflected before you walk away.
+5. When surprised, reconcile first
+   (`lab_registry.py reconcile`) before touching anything.
+
+VM lane: `labgrid/qemu-x86-64.yaml.example` in conwrt-bench is the
+pattern — the client runs ON ai-legion (QEMUDriver executes where the
+client runs), pristine per-acquire boots via `snapshot=on`. The on-target
+package harness is `tests/happy-path/run.sh --artifact <extracted pkg>`;
+the artifact is the published bytes from a kind-`1063` event
+(hash-pinned via its `x` tag) or, for pre-tag candidates, a local
+`scripts/build-sdk-package.sh` build whose sha256 is recorded in the
+evidence.
+
 ## Contributing process
 
 Follow [CONTRIBUTING.md](CONTRIBUTING.md). The parts agents most often