Confirmed-security quick-wins from a full re-audit (6-agent subsystem fan-out,
Codex-reviewed). Each item ships with a regression test proven to fail pre-fix.
- files: gate file CONTENT reads (download / edit / FileEditor::load) behind the
`operate` ability so a read-only viewer keeps LISTING/browse access but can no
longer pull file bytes over the server credential (root -> /etc/shadow, keys,
.env). Hide the download/edit/delete controls from viewers in the blade, and
basename() Files::open() for parity with the other path builders.
- fail2ban: reject a leading dash/dot in validJail(). A jail name like "-h" or
"--help" was parsed by fail2ban-client as an OPTION (argument-injection) rather
than a positional jail; internal dashes ("nginx-http-auth") still pass.
- bruteforce: release the atomically-claimed ban-audit dedup slot when the
AuditEvent write throws, so a transient DB failure no longer suppresses the
auth.ip_banned audit for the whole 60s window (the ban itself is unaffected).
- update path: verify_update_request now FAILS CLOSED when UPDATE_HMAC_KEY is
unset (was fail-open -> any ./run marker could drive a root update). Pin the
self-update `git pull` to the recorded CLUSEV_BUILD_BRANCH and refuse a diverged
checkout, so a `git checkout` in the tree cannot redirect the next root update.
- wireguard: normalise the DNS input to a single comma list and QUOTE the WG_DNS
assignment in wg.env; an unquoted space previously split the assignment and
broke the next `. "$WG_ENV"`.
13 new regression tests. 614 tests, Pint, shellcheck, Codex review all green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Security re-audit + adversarial red-team of the honeypot deception and
brute-force ban layer. Fixes (all with regression tests):
HoneypotController::trap
- BAN EVASION: keyed ban off getRealMethod() not isMethod('post') — the
latter honors _method / X-HTTP-Method-Override, so a POST spoofed as PUT
dodged the ban entirely. Now any non-safe verb to a decoy bans.
- REFLECTED-BAN vs EVASION: removed the Origin-based skip; the Origin header
is attacker-controlled, so it was a trivial universal ban-evasion (send a
foreign Origin) worse than the narrow griefing it prevented.
- UNMASK/500: mb_scrub all attacker-controlled meta (ua/query/path/creds) so
invalid UTF-8 can't throw JsonEncodingException on the array-cast column.
- Ban now runs FIRST and independently of the audit (separate guarded
try/catch) so a failing audit can never skip the ban; deceive() always 200.
- AUDIT DoS: per-IP hourly cap via atomic Cache::add(0)+increment.
DetectHoneytoken::trip
- Same guarded ban-first / audit-second structure + per-IP audit cap
(bounds an exempt or rotating source) + mb_scrub on the path.
BruteforceGuard::banNow
- Dedup the auth.ip_banned audit per IP+reason on a 60s window (was
full bantime): collapses the scanner flood without masking a legitimate
re-ban after an operator unban, and still audits distinct reasons.
Supply-chain / release / dashboard
- Pin prod base images (caddy/mariadb/redis) by @sha256 digest.
- Atomic branch+tag push in clusev-release.sh (no orphan untagged commit).
- Route github.ref_name through a validated $TAG env in ci-staging.yml.
- Strip inline credentials from the origin URL in set-repository-url.sh
(greedy match handles an @ inside the password).
- WireGuard peer name via Js::from in wire:click; settings badge shows the
real role; Threats pill/top_ip account for honeypot_login attempts.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Audit leak-hygiene findings:
- The operator's real internal dev-VM/WireGuard IPs (10.10.90.136/.162/.165) were shipped in
vite.config.js (hmr.host), lang hint/placeholder strings, a code comment and the CHANGELOG.
vite HMR host is now env-driven (VITE_HMR_HOST, dev .env only); all other occurrences are replaced
with RFC5737 documentation addresses (203.0.113.x). Tests updated in lockstep.
- docs/ is export-ignored from the git tree and named in promote.sh's refuse-to-publish list, but was
never .dockerignore'd, so `COPY . .` baked the private docs (with private host/URL refs) into the
public image, which the git-tree leak-guard never inspects. Added docs/ (and art/) to .dockerignore.
- /update.log (new repo-root update transcript) gitignored.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Security audit (12-angle adversarial workflow) findings on the container→host boundary:
- .env was created world-readable (cp of a 644 .env.example, never chmod'd) exposing the APP_KEY
that decrypts the whole fleet's SSH vault. Now chown clusev + chmod 0600 before any secret write.
- The root updater followed attacker-planted symlinks in the container-writable ./run (update-phase.json,
update.log) → arbitrary root file write. Status writes now go via a temp + rename(2) (never follow a
link); the update transcript moved to the repo root (not the ./run bind mount).
- The generated UPDATE_HMAC_KEY was dead. The app now HMAC-signs the update-request marker
(config/clusev.php update_hmac_key; DeploymentService) and watch.sh verifies it before running a
root update, so a stray/limited write to ./run can't drive one (a full container compromise holds
the key, so this is a bar-raise + integrity check; the standing guarantee is the marker only ever
re-installs the trusted remote's code, never attacker code).
- force_kv/set_kv wrote unescaped values into a sed replacement — a & | or newline in an operator's
domain/email/port could corrupt .env or inject a sed command. Values are now CR/LF-stripped and
sed-escaped, and the HTTP port is validated numeric/in-range.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A dashboard-triggered update that failed on the host left the /update-progress
page spinning to a vague 10-minute timeout — no error was ever shown, because
nothing wrote a failure marker for the page to read. Fix: make failures visible
and prevent the credential-prompt hang class.
- docker/restart-sentinel/watch.sh: run update.sh under `timeout -k 30 1800` with
its output tee'd to run/update.log; on ANY non-zero exit (update.sh or its
exec'd install.sh) — and on the early compose-missing / updater-missing paths —
write {"stage":"error"} to run/update-phase.json via write_update_error().
- update.sh: export GIT_TERMINAL_PROMPT=0 + GIT_HTTP_LOW_SPEED_* and wrap the pull
in `timeout 300`, so a private-repo credential miss fails fast instead of
hanging on a non-interactive prompt.
- DeploymentService::updatePhase(): whitelist the 'error' stage.
- update-progress.blade.php: add an error state (#js-status-error) + showError();
the status-feed poll now stops and shows it on stage=error, marking the active
phase red — instead of looping.
- lang/{en,de}/update.php: error_heading / error_hint / back_button.
- UpdateProgressTest: feed surfaces stage=error; page renders the error branch.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Reworks the Server-Details history graph from a static Livewire-rendered SVG into a
self-contained Alpine island fed by a JSON endpoint, addressing the feedback:
- Smooth Catmull-Rom curves instead of angular segments.
- Continuous line with an area gradient fill; the line breaks ONLY on a real outage
(delta-t > 2.5x the bucket), not on every empty bucket — no more dropouts.
- Hover crosshair + tooltip (time + CPU/MEM/DISK at the point).
- Instant client-side range switching (1h/24h/7d/30d) via
GET /servers/{server}/history.json — no Livewire round-trip.
- X-axis labels moved below the plot (no longer overlapping the 0 gridline).
MetricHistory::series() now returns only non-empty buckets plus from/now/bucket. The
chart wrapper is wire:ignore so the parent's 10s poll never clobbers the Alpine state;
the island self-refreshes every 60s. The old Livewire range logic is removed.
Codex review: clean (auth-gated route, range clamped, numeric-only JSON, guarded hover
bindings, R3/R4-compliant). Browser-verified: smooth curves, area fill, working tooltip
+ crosshair, instant range switch, zero console errors. 446 tests green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A CPU/RAM/Disk history graph over selectable ranges (1h/24h/7d/30d) on the server
detail page, backed by a persisted time series (the live 15s view stays cache-only).
- metric_samples table (server_id cascade, cpu/mem/disk %, load, sampled_at indexed).
- clusev:sample-metrics (scheduled every minute, mirrors clusev:wg-sample): persists
one row per server from the cached live reading, prunes beyond --retention (30d).
- MetricHistory::series() buckets + averages samples for a window; empty buckets stay
null so the chart shows gaps, not fabricated values.
- Servers/Show: range selector + a gap-aware SVG line chart (theme-token strokes,
R3-compliant, mirrors the dashboard chart). historyRange is clamped via setRange()
and the updatedHistoryRange() hook (it's a public, client-settable property).
Codex review: fixed a right-edge bucket off-by-one, clamped percentages on write, and
guarded the public range property against direct client values. Browser-verified: the
chart renders with data, range buttons, legend and gaps, zero console errors. 447
tests green, pint clean.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Verified live against a real Debian 12: the hardening actions work end-to-end
(root + password self-lockout guards refuse correctly, auto-updates and ufw enable,
disabling password auth persists and blocks password logins). One gap surfaced: on
a minimal image without rsyslog the distro-default fail2ban sshd jail (logpath
/var/log/auth.log) fails to start — "Have not found any log file". Pin the sshd jail
to `backend = systemd` (reads the journal) via a jail.d drop-in on enable:
log-source-agnostic and valid on every supported (systemd) target. fail2ban then
comes up active with the sshd jail reading the journal.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Disabling SSH root login (PermitRootLogin no) severs Clusev's OWN access when it
connects AS root — afterwards neither key nor password reaches root. The
password-auth toggle already guards against self-lockout; the root toggle did not.
Refuse to disable root login while the active credential's username is root, with a
clear message: add a non-root sudo user + key and switch Clusev to it first.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The self-update progress screen reset its elapsed timer to 0 and dropped the
phase checklist back to step one on every browser refresh — and the long image
build made it look like the update had restarted. Root causes: the timer and
phase state lived only in page JS (Date.now() at page load), and a
30s-before-page-load staleness guard rejected the real phase feed whenever a
stage had run longer than 30s (which the image build always does).
- DeploymentService: requestUpdate() now also writes a best-effort
run/update-started.json {at,from}. New updateStartedAt()/updatePhase() expose
the server-side start + current macro-stage; updatePhase() requires a positive
timestamp so a corrupt or stale 'done' can never drive a premature finish.
clearUpdateRequest() also removes the start marker.
- routes: /update-status.json delegates to updatePhase(); /update-progress
passes startedAt/initialPhase/hasFeed. A stage is trusted only when tied to
THIS update's start (startedAt) and written at/after it — a leftover stage from
a previous run can no longer resume or trigger a premature completion.
- update-progress.blade.php: the elapsed timer is anchored to the server-side
start (counts the real elapsed across a refresh); the checklist resumes at the
live stage; the live feed is authoritative (corrects any time-heuristic
overshoot) and the heuristic stays off once a real stage is known.
Verified: 28 UpdateProgressTest cases + full suite (426) green, pint clean, and a
headless browser load+reload mid-update keeps the timer counting and the phase in
place with zero console errors. Codex review: two premature-completion edge cases
found and fixed (timestamp validation + start-anchored freshness).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- PromotionService::publicTags() is called on every Release render incl. the 5s
page poll; the anonymous GitHub tag API is rate-limited (60/h per IP). Cache it
120s so the panel never gets rate-limited to empty. (Test flushes cache in setUp.)
- yank.yml: replace the loose shell-glob tag guard with an anchored grep regex so
a garbage/space/metachar tag dispatched straight from the GitHub UI is rejected
before reaching git push --delete (the panel already validates tag ∈ publicTags).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
'Staging' is not a single server: any beta-channel server pulls a release itself,
so a per-server 'deployed' step in the dev dashboard was arbitrary + contradicted
that model. The pipeline now tracks only the global build facts of a beta —
Tag (Gitea) → Mirror (GitHub-private) → CI. Removes PipelineStatus::testStep, the
test-server Setting/input/saveTestServer, and the now-orphaned lang keys.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- ReleaseChecker: mix a short hash of clusev.repository into the tag + changelog
cache keys so a cached value can never be reused across a host/repo change (the
promotion pipeline re-points CLUSEV_REPOSITORY between the private mirror and the
public repo).
- Strengthen the GitHub User-Agent test to assert the value (Clusev-Panel), not
mere presence — Guzzle always sets some UA, so the old check was a tautology.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
ReleaseChecker + the changelog fetch were hardwired to the Gitea API path, so a
GitHub-hosted public repo would never resolve a release. Centralise the repo-URL
parsing and branch by host: GitHub uses api.github.com for tags and
raw.githubusercontent.com for CHANGELOG.md (with the User-Agent GitHub requires);
Gitea keeps its /api/v1 paths. The changelog fetch moves into ReleaseChecker so
all host logic lives in one place.
Make the update-check tests hermetic: they now set clusev.repository themselves
(neutral host) instead of depending on the ambient .env — so they pass in CI with
an empty .env.example — and add GitHub-host coverage for tags + changelog.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Opening a binary / non-UTF-8 file in the file manager threw "undefined is not
valid JSON". The editor bound the raw file bytes to its public $content
property; Livewire JSON-encodes the component snapshot, and invalid UTF-8 makes
PHP's json_encode return false, so the client received an empty response and
JSON.parse("undefined") failed (the trace in mergeNewSnapshot/deepClone).
Fixed at both layers: FleetService::readFile blanks the content when it detects
binary/non-UTF-8 (returns the binary flag instead), and FileEditor::load only
binds content for non-binary, non-oversize files. Binary/oversize files show
their existing notice; the snapshot is always valid UTF-8.
Tests: load blanks binary content (keeps the snapshot valid) and keeps valid
text content. 374 pass, Pint clean.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two fixes the operator asked for:
- The "what's new" preview was empty. availableChangelog fetched CHANGELOG.md at
ref="0.9.50", but release tags are "v0.9.50" — the raw fetch 404'd and the
panel silently stayed empty. Use the v-prefixed tag as the ref. The existing
test masked it with a wildcard fake; tightened with Http::assertSent.
- The sidebar now shows a "1" badge on the Version item when an update is
available, so an update is visible on every page, not only the Versions page.
Driven by a new ReleaseChecker service: updateAvailable() is a pure cache read
(no network in the sidebar); a scheduled clusev:check-update (every 30 min,
anonymous/read-only) keeps the cache warm so the badge is accurate even before
the operator opens the Versions page.
Refactor: the remote tag lookup (fetch + newest-version) moved out of the
Versions component into ReleaseChecker; the component delegates. nav-item gains
an optional badge prop.
Tests: ReleaseChecker (cache-read update-available, refresh caches, command warms
cache, sidebar badge renders only when available) + the v-prefixed-ref assertion.
371 pass, Pint clean, /versions loads 200 with no console errors; badge + the
"What's new in v0.9.50" panel verified in the browser. Lang parity kept.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Three things the operator asked for on the WireGuard page:
- DNS is no longer hardcoded to 1.1.1.1. New host action set-dns (_set_dns +
cmd_set_dns + write-bridge branch + WgBridge validation, IPv4 list only),
stored as WG_DNS in wg.env (appended for pre-existing tunnels), surfaced via
the collector + WgStatus, baked into new peer configs (DNS = ${WG_DNS:-1.1.1.1}).
Editable in the Server tab; e.g. point clients at your own gateway.
- The page is split into two independent tabs: "Peers" (list / add / remove /
traffic) and "Server configuration" (endpoint / DNS / port / subnet / gate /
SSH lock + server identity). #[Url] $tab makes it deep-linkable (?tab=server).
- Offline detection: ONLINE_WITHIN 180s -> 150s (PersistentKeepalive=25 keeps
active peers re-handshaking well within it, so no flapping; a disconnect now
shows offline ~2.5 min sooner). Endpoint hint clarified: the port is optional
(auto-appended), no need to repeat the listen port.
Tests: set-dns (write+audit, reject), tab switch/clamp, dns in status; the two
configured-render tests now switch to the server tab for server content. 353
pass, shellcheck clean, Pint clean, both tabs load 200 with no console errors,
lang parity 89/89.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Verified dead-code + duplication cleanup over the WireGuard feature surface
(found via an audit workflow, every item adversarially re-verified by grep):
Dead code removed:
- lang keys not_configured_title / not_configured_body (both locales) — orphaned
when the unconfigured empty state became the in-page setup form (setup_title /
setup_intro). 0 references anywhere. de/en parity kept (83/83).
- WgTraffic::series() 'buckets' return key — written, never read by blade/route/test.
- blade $fmtB byte-formatter — byte-identical duplicate of the top-level $fmtBytes
already in scope; collapsed onto $fmtBytes.
DRY / readability (behavior-preserving):
- Index: PEER_NAME_RE constant (was the same regex literal at 5 sites), portInRange()
+ clampWindow() helpers (window default now self::WINDOWS[0]), openConfirm() helper
(5 confirm-modal openers shared one dispatch), onFlag() helper (2 apply handlers).
- WgBridge: validPort() helper (port check was duplicated verbatim).
- clusev-wg.sh: _write_server_conf() shared by both setup paths (the byte-identical
server bring-up was copy-pasted); single CIDR_RE constant for the 3 subnet checks;
fixed a stale header comment that still claimed SSH is "NEVER matched" (the v0.9.40
ssh-lock can match 22); documented the defence-in-depth re-validation in _set_*.
- WgStatus: documented server.pubkey is part of the /wg-status.json contract (kept).
No DB/migration changes (peer_name write-only column left intact — needs a
coordinated migration). 348 tests pass, shellcheck clean, Pint clean, R12 verified
in both the configured + unconfigured states (200, no console errors, no leaked
keys, traffic formatter renders 1.5 MB / 200 KB correctly).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two dashboard additions to the WireGuard page:
- Peer config download: the show-once new-client view gets a "Download"
button next to the QR code (streams clusev-wireguard.conf) plus an
inline hint explaining QR-on-phone vs download-on-PC import.
- SSH lock (port 22): an optional toggle in WireGuard settings that walls
off host SSH to the tunnel via a dedicated CLUSEV-WG-SSH chain on the
host INPUT chain. Fail-open by construction: ESTABLISHED,RELATED + lo +
wg0 + WG-subnet RETURN before the DROP (live sessions survive, SSH over
the tunnel stays open); _ssh_gate_on refuses while wg0 is down; the boot
unit's ConditionPathExists=wg0 and an in-function self-guard keep a
broken tunnel from ever applying the DROP; ssh_gate_apply re-validates
the subnet shape and fails open on a corrupt wg.env. Strong lock-out
warning shown inline AND in the confirm modal (recommend a backup peer).
New CLI escapes: clusev wg ssh-lock / ssh-unlock; clusev wg down clears
both gates. Panel gate and SSH gate are independent toggles.
Write-bridge: new gate-ssh-on/gate-ssh-off actions (no args, host
whitelist + ConfirmToken wgSshGate, single-use uid-bound). Status collector
now reports gate.ssh. Adversarial lock-out review: all 8 failure modes
refuted. 348 tests pass, shellcheck clean, Pint clean, R12 verified.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Even with authenticatorAttachment=cross-platform, Bitwarden's extension
intercepted navigator.credentials.create and offered to save a passkey. Add the
WebAuthn L3 'hints: [security-key]' to both registration and assertion options;
browsers and passkey managers that honor hints step aside and let the hardware
key prompt through. Injected into the client payload only (the session copy used
for server validation is unchanged).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Registration set no authenticatorSelection, so the browser offered the
passkey-manager 'save passkey' flow (Bitwarden) instead of prompting for the
YubiKey. Now request a cross-platform, non-resident, presence-only credential:
authenticatorAttachment=cross-platform (roaming key, not a platform/manager
passkey), residentKey=discouraged (non-discoverable second factor — what keeps
passkey managers out), userVerification=discouraged (single touch, no PIN).
Login assertion also uses discouraged so it is: insert key, one tap, done.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
WebauthnService::available() keyed on request()->isSecure(), which is false in
external-proxy mode (the forwarded https scheme isn't trusted without
TRUSTED_PROXY_CIDR) — so hardware keys were locked even though the panel runs
over HTTPS on the domain. Gate on being reached via the active domain instead
(always HTTPS through the front door); the bare-IP recovery host stays locked.
A core security feature no longer depends on an optional proxy setting.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- reverbClient() now returns wss://<domain>:443 when the active domain is being
served (the HTTPS front door), instead of deriving wss://host:80 from a request
that looks like plain HTTP:80 behind an external TLS proxy. Bare-IP recovery
stays plain ws.
- Caddy no longer redirects the /app/* /apps/* WS paths to https when the upstream
already terminated TLS (X-Forwarded-Proto: https) — that redirect broke the
WebSocket in external-proxy mode.
- The 'TLS active' status card no longer claims Let's Encrypt issuance in
external-proxy mode; it states TLS is provided by the upstream proxy.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Mirror the update button (0.9.15): requestRestart() now returns bool and
restartNow() shows an error toast (pointing at sudo ./update.sh) when the
sentinel can't be written, instead of a 'restarting' state that never resolves.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The prod image baked the app user at uid 1002, but install.sh chowns the
bind-mounted ./run to the clusev host user. On a host where clusev != 1002 the
app process could not write the restart/update sentinel — the dashboard button
showed 'läuft' while nothing happened on the server. Pass APP_UID/APP_GID =
HOST_UID/HOST_GID into the prod build so the in-image app user matches clusev
and ./run is writable. Also surface a failed sentinel write as an error toast
instead of a stuck 'running' state (requestUpdate now returns bool).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add a host-watcher-backed update button on Version & Releases. The container
(no Docker socket) writes an update sentinel; a new root-run systemd unit
(clusev-update.path/.service) runs watch.sh update -> update.sh (git pull +
idempotent install). The sentinel is consumed before running so a persistent
failure can't loop the path unit.
The button only appears once a check finds a newer release, is per-user
throttled (3/10min, auto-expiring) and audited (deploy.update_request); the UI
warns the dashboard is briefly down during the rebuild. install.sh now installs
both the restart and update host units (update unit runs as root for
docker/systemd/apt).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
For the common case where a domain was configured before its DNS pointed here, the
operator can now request the Let's Encrypt cert from System -> Domain & TLS instead of
waiting for the first HTTPS visitor:
- DeploymentService::requestCertificate() does a DNS pre-check (domainResolvesHere via
dns_get_record vs serverPublicIp) and, only when DNS is NOT a clear mismatch, triggers
Caddy's on-demand issuance with an internal handshake (probeCertificate:
`curl --connect-to <domain>:443:caddy:443` — the dial target is pinned to the caddy
service, the domain only sets SNI/Host, so no SSRF). A clear mismatch returns early and
never calls ACME (Let's Encrypt rate-limit protection). The outcome is persisted to the
`tls_cert_status` Setting so the page shows a status without an ACME-triggering probe on
every load.
- System\Index::requestCertificate() is per-user throttled (cert-request:<id>, 5/10min,
auto-expiring -> never a control-plane lockout) and audited (tls.cert_request). The
status line + button render only in caddy mode with an active domain; external mode
shows a note; bare IP hides it.
Automatic on-demand still works on the first HTTPS handshake regardless — this only adds
explicit control + a status. Tests cover the no-handshake-on-mismatch guard, the
issued/failed/not-applicable states, and the per-user throttle. Full suite 192 green;
Codex clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
boban