docs: cmake/platform.cmake belongs to PKG/QA, not PORT (ADR 0020 amendment)

As written, ADR 0020 gave PORT ownership of cmake/platform*.cmake while also
making PORT wait for Phase 0 — but Phase 0's seams are exactly what needs
velox_platform_sources() to select platform/linux/*.cpp. PORT cannot start until
Phase 0 lands, and Phase 0 cannot land without the file PORT owns. PKG/QA caught
the circularity before writing anything, which was the right call.

Central OS detection is root build infrastructure, so it stays with PKG/QA and
lands during Phase 0. PORT keeps the per-OS backend directories and the macOS
and Windows packaging, and owns no part of cmake/.

Also drops a hardcoded test count from the PORT definition of done; the suite
number moves every round and the gate is that it stays green.

Co-Authored-By: Claude Opus 5 <[email protected]>

CLAUDE.md

@@ -12,6 +12,7 @@ Read this before touching anything. Then read your lane brief in [docs/agents/](
 | GUI | `gui/` | everything else |
 | EXT | `extension/` | everything else |
 | PKG/QA | `packaging/`, `.github/`, `tools/testserver/`, `tests/integration/`, `tests/e2e/`, root build files | any lane's feature code |
+| PORT | `**/platform/<os>/**`, `packaging/macos/`, `packaging/windows/` | all logic, and `cmake/` — see [ADR 0020](docs/adr/0020-cross-platform-strategy.md) |
 
 If your task seems to require editing another lane's files, that is a signal the interface
 is wrong. **File the request; don't reach across.**

cli/man/velox.1

@@ -0,0 +1,129 @@
+.TH VELOX 1 "2026-09-12" "Velox" "User Commands"
+.SH NAME
+velox \- command-line client for veloxd, the Velox download manager daemon
+.SH SYNOPSIS
+.B velox
+.I command
+.RI [ options ]
+.SH DESCRIPTION
+.B velox
+talks to
+\fBveloxd\fR
+over its Unix-domain socket and gives a scriptable, terminal-first view of the same
+downloads the GUI and the Firefox extension see. It does no downloading itself \(em every
+command is a thin JSON-RPC call; the daemon owns all state.
+.PP
+.B veloxd
+must already be running (see
+.B ENVIRONMENT
+below for how
+.B velox
+finds it, and
+.BR systemctl (1)
+.RI ( "systemctl --user status velox" )
+for whether it's up). If nothing is listening on the socket,
+.B velox
+exits with status 3 rather than hanging.
+.SH COMMANDS
+.TP
+.BI "add " url " [\-\-dir " dir "] [\-\-out " name "] [\-\-segments " n ]
+Add a new download. Prints the new task's id and initial state.
+.RS
+.TP
+.BI "\-\-dir " dir
+Destination directory. Must resolve inside one of the daemon's configured
+.B saveTo.allowedRoots
+or the call fails; omit to use the configured default download directory.
+.TP
+.BI "\-\-out " name
+Filename to save as. Omit to derive one from the URL (or, once the daemon has probed it,
+from the server's own
+.BR Content-Disposition ).
+.TP
+.BI "\-\-segments " n
+Requested connection count for this download, 1\(en32. The daemon may use fewer \(em a
+per-host cap or a source that turns out not to support resuming both lower it. The
+.B ls
+table (and
+.RI "\-\-json's " segments
+field) show what was actually granted, not what was asked for.
+.RE
+.TP
+.B ls
+List every download the daemon knows about: id, state, progress, and filename. With
+.BR \-\-json ", the raw " download.list " result instead of the table."
+.TP
+.BI "pause " id " [" id " ...]"
+Pause one or more downloads by id. A download already paused, or already finished, is
+left alone \(em not an error.
+.TP
+.BI "resume " id " [" id " ...]"
+Resume one or more paused downloads.
+.TP
+.BI "rm " id " [" id " ...] " "[\-\-delete\-file]"
+Remove one or more downloads from the list. Without
+.BR \-\-delete\-file ,
+any partial data on disk
+.RI ( .veloxpart / .veloxpart.meta )
+is discarded but a
+.B completed
+file is left in place. With
+.BR \-\-delete\-file ,
+the finished file is deleted too \(em there is deliberately no default for this flag; you
+must say which you mean every time.
+.SH OPTIONS
+.TP
+.B \-\-json
+Print the raw JSON-RPC result instead of a formatted table. Works with every command;
+combine with
+.BR jq (1)
+for scripting. On error, the JSON form is an
+.B {"error": {...}}
+object on stdout rather than a message on stderr.
+.TP
+.B \-h ", " \-\-help
+Print usage and exit 0.
+.SH EXIT STATUS
+.TP
+.B 0
+Success.
+.TP
+.B 1
+The daemon reached the call but returned a JSON-RPC error (bad task id, path outside the
+allowed roots, and so on). The message is on stderr, or in the JSON error object with
+.BR \-\-json .
+.TP
+.B 2
+Usage error \(em missing argument, unknown option, or unknown command.
+.TP
+.B 3
+Could not reach
+.B veloxd
+at all: not running, or its socket is missing or unreachable.
+.SH ENVIRONMENT
+.TP
+.B XDG_RUNTIME_DIR
+.B velox
+connects to
+.IR "$XDG_RUNTIME_DIR/velox/velox.sock" .
+If unset, it falls back to
+.IR /run/user/ <uid> ,
+matching
+\fBveloxd\fR's
+own resolution \(em the two must agree for the client to find the daemon, so this is
+normally left to the desktop session's default rather than set by hand.
+.SH FILES
+.TP
+.I $XDG_RUNTIME_DIR/velox/velox.sock
+The daemon's Unix-domain socket, mode 0600, same-UID only (\fBSO_PEERCRED\fR checked on
+every connection \(em this is the transport's authorization, not an extra login).
+.SH SEE ALSO
+.BR systemctl (1),
+.BR jq (1)
+.PP
+.I docs/01-architecture.md
+and
+.I docs/agents/AGENT-DAEMON.md
+in the Velox source tree for the daemon's own build order and the wire protocol
+.B velox
+speaks.

daemon/CMakeLists.txt

@@ -75,6 +75,7 @@ target_link_libraries(veloxd_sched PUBLIC velox::proto velox::core veloxd_store
 add_library(veloxd_rpc STATIC
     src/rpc/runtime_dir.cpp
     src/rpc/single_instance.cpp
+    src/rpc/systemd_activation.cpp
     src/rpc/event_loop.cpp
     src/rpc/event_hub.cpp
     src/rpc/uds_server.cpp

daemon/docs/deferrals.md

@@ -7,7 +7,7 @@ close. Kept here (not buried in commit messages) so the next pass can see them a
 |---|---|---|---|---|
 | ~~D7~~ | **Closed — `capture.offer` is real.** Applies `capture.enabled`/`excludedHosts`/`monitoredExtensions`/`monitoredMimeTypes`/`minSizeBytes` from settings, then the rules table (`store::Rules` + CORE's `vdm::rules::match_rules`/`glob_match` — DAEMON only converts its own stored `proto::Rule` JSON into CORE's plain `vdm::rules::Rule` vocabulary, per that header's own layering note), resolves the category folder (a rule's explicit `categoryId`/`saveDir`, else `store::Categories::guess_by_extension` — the same extension-guess `download.probe`'s `suggestedCategoryId` already used, now shared instead of duplicated), dedupes against active (non-terminal) tasks by exact URL, and on `take` calls `add_one()` — the same path `download.add` itself uses — so a captured download is a real, admitted, persisted task, not a special case. The 750 ms deadline (CLAUDE.md §4 / AGENT-DAEMON.md build step 6) is checked cooperatively between every step via a new `rpc::CaptureDataSource` seam (real impl wraps `store::*`; a test fake can jump its own clock forward to simulate "the store was slow just now" with zero real sleep) — catches the realistic failure mode (several slow steps adding up) though it can't preempt one pathologically stuck single call. Verified against real `veloxd` + `tools/testserver`: a monitored-type offer answers in ~5ms and actually creates + downloads the task; an unmonitored type, an excluded host, a rule-vetoed host, and a second offer for a still-active URL all answer `ignore` with the right `reason`; a bad category save dir surfaces its real `-32011` rather than being swallowed. New `capture_offer_test` covers all of the above plus the deadline itself (two cases, one per "slow" checkpoint), asserting real wall-clock time barely moves even though the fake clock jumped 2 simulated seconds — proof the check reads the injected clock, not a disguised sleep. | `rpc/capture_data_source.hpp`, `rpc/dispatcher.{hpp,cpp}`, `store/rules.{hpp,cpp}`, `store/categories.{hpp,cpp}`, `store/tasks.{hpp,cpp}` | — | done |
 | ~~D8~~ | **Closed alongside D7** — `capture.getRules` returns the same settings-backed `enabled`/`monitoredExtensions`/`monitoredMimeTypes`/`minSizeBytes`/`excludedHosts`/`bypassModifier` capture.offer itself reads, so the two can never drift. `rulesVersion` is a constant `1` — there is no persisted revision counter yet (nothing writes `rules.*` outside this process's own lifetime to need one across a restart), and the extension already re-fetches on `event.settings.changed` regardless of what this number does; noted in case a real counter becomes worth adding later. | `rpc/dispatcher.cpp` | `rulesVersion` is a placeholder constant | — |
-| D1 | Pairing prompt is `EnvAutoApprover` (needs `VELOX_PAIR_AUTO=1`) | `rpc/pairing.hpp`, `main.cpp` | A GUI dialog / `org.freedesktop.Notifications` approver is integration work | Build step 7 (systemd + notifications) |
+| D1 | Pairing prompt is `EnvAutoApprover` (needs `VELOX_PAIR_AUTO=1`) | `rpc/pairing.hpp`, `main.cpp` | A GUI dialog / `org.freedesktop.Notifications` approver is integration work | the `org.freedesktop.Notifications` half of build step 7 — the systemd half closed as D11 below |
 | — | **D1, checked this pass, not attempted:** `libdbus-1-dev` (or `libsystemd-dev` for `sd-bus`) has no headers installed in this build environment — only the runtime `.so`s (`dpkg -l`/`apt-cache policy` confirm `libdbus-1-3` present, `libdbus-1-dev` not, "Candidate" available but not installed). A real notification-backed approver needs one of those linked into `veloxd`, which is a new build dependency for `daemon/CMakeLists.txt` (`find_package`/`pkg_check_modules`) and — since packaging manifests need to know about it too — arguably a decision to surface rather than something to reach for silently mid-session. `PairingApprover::approve()` is also still synchronous by shape (its own doc comment already says so: "the real notification-backed approver will run async and is not this shape") — swapping it for the async pattern this session built for `download.probe` (`rpc::TaskActionPort` + the server-layer deferred-reply special-case) is the right shape once there's a real implementation to justify the churn; reshaping the interface with nothing behind it yet would just be churn. Left `EnvAutoApprover` in place rather than build a fragile hand-rolled D-Bus wire client to avoid the missing headers — a broken pairing approver is worse than an honest stub. | `rpc/pairing.hpp` | missing dev headers + an undiscussed new dependency | once `libdbus-1-dev`/`libsystemd-dev` is available and the dependency is approved |
 | ~~D2~~ | **Closed** — `download.probe` is real on both transports. It's genuinely async (the engine's probe pool, up to the schema's 30s `x-deadlineMs`) and so cannot fit `VeloxDispatcher::on_download_probe`'s synchronous `HandlerResult<T>` return — `uds_server.cpp`/`ws_server.cpp` special-case `"download.probe"` before the generic `dispatch()`, exactly the way they already special-case `session.hello`/`session.subscribe`, and queue the reply whenever the callback fires. `rpc::TaskActionPort::probe_now` (kept in proto/std terms, no `vdm::net::*`, so `veloxd_rpc` never needs `core/include`'s vdm headers) is what both transports call; `sched::Scheduler::probe_now` is the implementation — builds a `vdm::net::ProbeRequest`, runs it on the engine's probe pool, maps a failure to `-32013 ProbeFailed` (with `data.httpStatus` when there was one), and fills `suggestedCategoryId`/`suggestedSaveDir` with a plain extension match against the categories table (not the real rules engine — that's still D3). Verified live: a real probe answers in ~5ms; a bad host maps to `-32013`; a connection issuing a 10s `slow-loris` probe does not block a second connection's `download.list` (answered in ~1ms) — confirms the async design actually keeps the loop free, not just compiles. | `rpc/task_action_port.hpp`, `rpc/{uds_server,ws_server}.{hpp,cpp}`, `sched/scheduler.{cpp,hpp}` | — | done |
 | D3 | Stub handlers for the rest: `grabber.*`, `media.*` | `rpc/dispatcher.cpp` | HLS/DASH grabber and media-variant support don't exist anywhere in this build yet — a bigger feature than a store-wiring pass | M4 territory, per AGENT-DAEMON.md |
@@ -26,4 +26,5 @@ close. Kept here (not buried in commit messages) so the next pass can see them a
 | ~~D4b~~ | **Closed** — `download.pause`/`resume`/`start`/`cancel` and `queue.start`/`stop` all drive the scheduler now, and apply *immediately* (not deferred to the next tick — pausing/resuming/cancelling a live transfer can't wait up to 1s, and per ADR 0013 §3 the governor never touches a user-owned pause on its own). New `rpc::TaskActionPort` interface (owned by `rpc/`, implemented by `sched::Scheduler`) is the seam dispatcher.hpp depends on instead of `sched/scheduler.hpp` directly — avoids a real `veloxd_rpc` <-> `veloxd_sched` circular library dependency (`veloxd_sched` already links `veloxd_rpc` for `EventHub`). `Scheduler::user_pause/resume/start/cancel` + `pause_queue` engine-call-then-eager-transition, matching `tick()`'s existing `to_pause` pattern. Fixed a real bug hit while building this: `transition()` always overwrote `pause_reason` to NULL when the engine's own delayed pause-ack callback arrived with no explicit reason, clobbering whatever the actual initiator (user or governor) had just written — now it preserves the stored reason when none is supplied. Verified against real `veloxd` + `tools/testserver`: pausing a live single-segment throttled transfer freezes `downloadedBytes`, resume continues it from that point, cancel stops it; `queue.stop(pauseRunning:true)` pauses the queue's running task immediately. NOTE: `download.start`'s contract "a task in 'queued' jumps its queue" (priority bump) is not implemented — admission is still plain FIFO by `created_at`. | `sched/scheduler.{cpp,hpp}`, `rpc/task_action_port.hpp`, `rpc/dispatcher.{hpp,cpp}`, `store/queues.{cpp,hpp}` | — | done, except the queue-jump priority bump noted above |
 | ~~D5~~ | **Mostly closed** — `rpc/event_hub` fans out per-subscription; `session.subscribe` on both transports registers/updates/tears down a real subscription; `Scheduler::transition()` publishes `event.task.state` (with `previousState`) on every state change, scheduler-driven or engine-reported; `dispatcher::on_download_add` publishes `event.task.added`; a 250 ms timer batches `Scheduler::progress_snapshot()` into one `event.task.progress` array per AGENT-DAEMON.md item 5 / the schema's `x-maxRateHz: 4`. Verified live end to end. | — | `event.task.removed` has no source yet (`download.remove` is D3); `event.speed.global`, `event.notify`, `event.auth.required`, `event.settings.changed`, `event.grabber.progress` are unpublished — each lands with its owning handler | as each owning D3 handler lands |
 | ~~D6~~ | **Closed** — engine numbers now reach the store: `Scheduler::tick()` probes (`EnginePort::probe`) before every `start()`, persisting `sizeBytes`/`resumable`/validators via `Tasks::set_probe_result` before a byte moves; `Scheduler::persist_progress()` (called from `progress_snapshot()` *and* once more from `on_engine_state` right before `release()`/unmap on every terminal transition) writes `downloadedBytes`/`speedBps`/`segments`/`segmentDetail` from the engine's `Progress`, so a task that finishes between two 250 ms ticks (the common case for anything small or fast) still leaves real numbers instead of the pre-persistence defaults. `TaskSummary.segments` is sourced from `segments.size()` when the task has any (matching what actually lands in `segmentDetail`, per the schema's "exactly `segments` entries"), falling back to the engine's `effective_segments` (budget slots *held*, not necessarily physical range count — see `core/include/vdm/task/download.hpp`'s `Progress` comment) only pre-segmentation. `Tasks::set_final_bytes` tops up `on_finished`'s byte count as a last-resort backstop. Migration `0002` adds `speed_bps` to both `tasks` and `segments`, and fixes `segments.state`'s CHECK to include `'pending'` (0001 omitted it, so a pre-connect snapshot could never be written). Verified against real `veloxd` + `tools/testserver` (not just unit tests): `download.list`/`download.get` correct immediately after completion and after a daemon restart. | `sched/scheduler.{cpp,hpp}`, `store/{tasks,segments}.{cpp,hpp}`, `store/migrations/0002_*.sql` | — | done |
+| ~~D11~~ | **Closed — build order items 7 (the systemd half) and 9: `velox-nmhost`, socket activation, the systemd user units, and `velox(1)`.** `nmhost/src/main.cpp` (185 lines): a `poll()`-driven byte pump between Firefox's native-messaging framing on stdio (4-byte native-byte-order length prefix) and `veloxd`'s own NDJSON framing on the Unix socket — reframes each direction, no JSON parsing, no retry/backoff, exits the moment either side closes. Deliberately dependency-free (no `veloxd_*` library, no `nlohmann_json`) since it runs unconfined outside Firefox's sandbox whatever the packaging format. Two real bugs found and fixed while getting the integration test to actually pass rather than hang: (1) never set the pumped fds non-blocking, so the "drain what's available" read loop blocked on its own second `read()` instead of returning to `poll()`; (2) stdin and stdout are two different descriptors (0 and 1), not one — an early draft polled `POLLOUT` on fd 0, which is opened read-only, so EOF/writability were never both observable through the same `pollfd` entry. Both are exactly the class of bug a "trivial pump" invites and unit tests over the real binary (not just its helper functions) exist specifically to catch. `packaging/nativehost/com.velox.host.json` + its own `README.md` supersede `AGENT-DAEMON.md`'s stale "four locations" (spike S1 / ADR 0003 found only three are real — the fourth, `~/snap/firefox/common/.mozilla/...`, is not read by snap Firefox at all) and spell out the per-user-manifest / postinst implication for PKG/QA. `EnginePort`-style: `rpc/systemd_activation.cpp` is a from-scratch `sd_listen_fds()` (env vars only, no `libsystemd` link — `LISTEN_PID`/`LISTEN_FDS`, fd 3) that `UdsServer::start()` checks first, skipping its own create/bind/chmod/listen when systemd already bound the socket; `packaging/systemd/velox.socket` + `velox.service` are the unit pair, verified both by `systemd-analyze verify` and by an actual fork/dup2/execve simulation of the activation handshake (a real `session.hello` round-tripped over the handed-off fd with no `bind()` ever called inside the daemon for that run). `velox.service` deliberately skips `ProtectSystem=`/`ProtectHome=`/`ReadWritePaths=` — `saveTo.allowedRoots` is user-configurable to anywhere on the filesystem, and a sandbox here would turn a legitimately-configured save location into an opaque `EROFS`/`EACCES` instead of the daemon's own clear `-32011`. `cli/man/velox.1` documents the CLI as it actually exists today (`add`/`ls`/`pause`/`resume`/`rm`, `--json`, the three-tier `queue`/`settings` subcommands `AGENT-DAEMON.md` build step 8 originally sketched are not implemented in `cli/src/main.cpp` yet, so the page doesn't claim they are) — checked warning-free with `groff -mandoc -ww -z`. | `nmhost/{CMakeLists.txt,src/main.cpp,tests/}`, `daemon/src/rpc/{systemd_activation.{hpp,cpp},uds_server.cpp}`, `packaging/{nativehost,systemd}/`, `cli/man/velox.1` | — | done |
 | — | ~~Observed, not fixed (CORE, not this lane)~~ — **routed to CORE by the user.** `vdm::task::Progress.speed_bps` reads back as `0` for the whole lifetime of a live, real (non-fake) throttled download, despite `downloadedBytes` visibly advancing between polls — `core/src/task/download_task.cpp`'s per-worker EWMA never seems to produce a nonzero aggregate in this build. DAEMON passes `EnginePort::progress()`'s `speed_bps` straight through (`Scheduler::persist_progress`); nothing in this lane drops it. Still reproduces in the D4b live checks above (0 throughout a paused/resumed/cancelled transfer whose `downloadedBytes` visibly moved) — not re-filed, since it's already CORE's. |

daemon/src/rpc/systemd_activation.cpp

@@ -0,0 +1,37 @@
+#include "rpc/systemd_activation.hpp"
+
+#include <unistd.h>
+
+#include <cstdlib>
+#include <string>
+
+namespace velox::daemon::rpc {
+
+namespace {
+constexpr int kListenFdsStart = 3;  // SD_LISTEN_FDS_START
+}  // namespace
+
+int systemd_activated_fd() {
+    const char* pid_env = std::getenv("LISTEN_PID");
+    const char* fds_env = std::getenv("LISTEN_FDS");
+    int fd = -1;
+
+    if (pid_env != nullptr && fds_env != nullptr) {
+        try {
+            if (std::stol(pid_env) == static_cast<long>(::getpid()) && std::stol(fds_env) == 1) {
+                fd = kListenFdsStart;
+            }
+        } catch (...) {
+            // Malformed env from something other than systemd; treat as not activated.
+        }
+    }
+
+    // Contract: consumed once, then cleared, so a value meant for veloxd is never
+    // mistaken for one meant for a process it might itself exec later.
+    ::unsetenv("LISTEN_PID");
+    ::unsetenv("LISTEN_FDS");
+    ::unsetenv("LISTEN_FDNAMES");
+    return fd;
+}
+
+}  // namespace velox::daemon::rpc

daemon/src/rpc/systemd_activation.hpp

@@ -0,0 +1,20 @@
+#pragma once
+
+// Minimal sd_listen_fds(3) reimplementation — one function, no libsystemd dependency, for
+// the one fd velox.socket ever hands us. See velox.socket / velox.service in
+// packaging/nativehost's systemd unit pair: the socket unit binds
+// $XDG_RUNTIME_DIR/velox/velox.sock itself (before veloxd ever runs, so the very first
+// connection attempt after boot is queued by the kernel rather than refused) and execs
+// veloxd with that listening fd already open at fd 3, LISTEN_FDS=1, LISTEN_PID=<our pid>.
+
+namespace velox::daemon::rpc {
+
+// The systemd-activated listening socket fd, or -1 if this process was not socket-
+// activated (LISTEN_PID doesn't match our pid, or LISTEN_FDS is unset/not exactly 1 — more
+// than one would mean a unit file mismatch, since veloxd only ever asks for one socket).
+// Clears LISTEN_PID/LISTEN_FDS from the environment on the way out either way, per
+// sd_listen_fds's own contract, so a value meant for us is never mistaken for one meant for
+// a process veloxd might itself exec later.
+int systemd_activated_fd();
+
+}  // namespace velox::daemon::rpc

daemon/src/rpc/uds_server.cpp

@@ -12,7 +12,10 @@
 
 #include <nlohmann/json.hpp>
 
+#include <fcntl.h>
+
 #include "rpc/event_loop.hpp"
+#include "rpc/systemd_activation.hpp"
 #include "version.hpp"
 
 namespace velox::daemon::rpc {
@@ -74,6 +77,20 @@ UdsServer::~UdsServer() {
 }
 
 std::error_code UdsServer::start() {
+    // velox.socket (systemd user unit, socket activation): the unit binds this path itself
+    // before veloxd ever runs and hands the already-listening fd over at fd 3 — the first
+    // connection after boot is queued by the kernel rather than refused, and there is no
+    // window where a client sees ECONNREFUSED while the daemon is still starting. Skips
+    // create/bind/chmod/listen entirely; the socket file's lifecycle (including removal on
+    // stop) belongs to the unit, not to us, so bound_ stays false.
+    if (const int activated = systemd_activated_fd(); activated >= 0) {
+        ::fcntl(activated, F_SETFL, O_NONBLOCK);
+        ::fcntl(activated, F_SETFD, FD_CLOEXEC);
+        listen_fd_ = activated;
+        loop_.add_fd(listen_fd_, kRead, [this](int, unsigned) { on_listener_readable(); });
+        return {};
+    }
+
     if (path_.size() + 1 > sizeof(sockaddr_un::sun_path)) return errc(ENAMETOOLONG);
 
     const int fd = ::socket(AF_UNIX, SOCK_STREAM | SOCK_NONBLOCK | SOCK_CLOEXEC, 0);

daemon/tests/CMakeLists.txt

@@ -24,6 +24,7 @@ veloxd_test(sched_scheduler  LIBS veloxd_sched veloxd_rpc)
 veloxd_test(event_hub               LIBS veloxd_rpc)
 veloxd_test(store_categories_queues LIBS veloxd_store)
 veloxd_test(single_instance          LIBS veloxd_rpc)
+veloxd_test(systemd_activation        LIBS veloxd_rpc)
 veloxd_test(dispatcher_settings      LIBS veloxd_rpc veloxd_store)
 veloxd_test(capture_offer            LIBS veloxd_rpc veloxd_store)
 veloxd_test(dispatcher_misc          LIBS veloxd_rpc veloxd_store)

daemon/tests/systemd_activation_test.cpp

@@ -0,0 +1,53 @@
+// systemd_activated_fd(): the LISTEN_PID/LISTEN_FDS contract, without a real systemd.
+
+#include <unistd.h>
+
+#include <cstdlib>
+#include <string>
+
+#include "check.hpp"
+#include "rpc/systemd_activation.hpp"
+
+using namespace velox::daemon::rpc;
+
+namespace {
+void set_env(const char* k, const std::string& v) { ::setenv(k, v.c_str(), 1); }
+}  // namespace
+
+void run() {
+    // Not activated: neither var set.
+    ::unsetenv("LISTEN_PID");
+    ::unsetenv("LISTEN_FDS");
+    CHECK_EQ(systemd_activated_fd(), -1);
+
+    // LISTEN_PID for a different process: not us, so not activated.
+    set_env("LISTEN_PID", std::to_string(::getpid() + 1));
+    set_env("LISTEN_FDS", "1");
+    CHECK_EQ(systemd_activated_fd(), -1);
+    // Consumed regardless of the outcome — a stale value from some other process's
+    // exec chain must not leak into what veloxd checks next time.
+    CHECK(::getenv("LISTEN_PID") == nullptr);
+    CHECK(::getenv("LISTEN_FDS") == nullptr);
+
+    // Our own pid, LISTEN_FDS=1: activated, fd 3 (SD_LISTEN_FDS_START).
+    set_env("LISTEN_PID", std::to_string(::getpid()));
+    set_env("LISTEN_FDS", "1");
+    CHECK_EQ(systemd_activated_fd(), 3);
+    CHECK(::getenv("LISTEN_PID") == nullptr);
+
+    // Our own pid but LISTEN_FDS=2: a unit file mismatch (veloxd only ever asks for one
+    // socket) — refuse rather than guess which of two fds is the right one.
+    set_env("LISTEN_PID", std::to_string(::getpid()));
+    set_env("LISTEN_FDS", "2");
+    CHECK_EQ(systemd_activated_fd(), -1);
+
+    // Garbage LISTEN_FDS: not activated, not a crash.
+    set_env("LISTEN_PID", std::to_string(::getpid()));
+    set_env("LISTEN_FDS", "not-a-number");
+    CHECK_EQ(systemd_activated_fd(), -1);
+
+    ::unsetenv("LISTEN_PID");
+    ::unsetenv("LISTEN_FDS");
+}
+
+TEST_MAIN()

docs/08-porting.md

@@ -0,0 +1,85 @@
+# 08 — Porting guide (macOS first, Windows later)
+
+Read `docs/adr/0020-cross-platform-strategy.md` first; this file is the concrete inventory
+it refers to. **Rule: Linux behaviour never changes. `#ifdef` never appears in logic.**
+
+## Directory layout
+
+Platform backends live beside the code they serve, one directory per OS:
+
+```
+core/src/io/platform/linux/file_ops.cpp        core/src/io/platform/macos/file_ops.cpp
+core/src/io/platform/file_ops.hpp              # the seam — no #ifdef, no OS types
+daemon/src/rpc/platform/linux/{wakeup,peercred,instance_lock,runtime_dir}.cpp
+daemon/src/rpc/platform/macos/{...}.cpp
+daemon/src/rpc/platform/*.hpp                  # the seams
+cmake/platform.cmake                           # VELOX_OS_* + source selection
+```
+
+`cmake/platform.cmake` sets exactly one of `VELOX_OS_LINUX`, `VELOX_OS_MACOS`,
+`VELOX_OS_WINDOWS`, and exposes `velox_platform_sources(<target> <dir>)` which adds
+`<dir>/platform/<os>/*.cpp`. Selection happens there and nowhere else.
+
+## The seams
+
+Five interfaces cover the whole port. Signatures are indicative, not binding — the owning
+lane settles them in Phase 0.
+
+| Seam | Interface | Why |
+|---|---|---|
+| File preallocation | `Result<void> preallocate(int fd, uint64_t bytes)` | `posix_fallocate` is Linux/glibc |
+| Cache advice | `void advise_dontneed(int fd, uint64_t off, uint64_t len)` | `posix_fadvise` has no macOS equivalent |
+| Durable flush | `Result<void> flush_durable(int fd)` | `fdatasync` vs `F_FULLFSYNC` |
+| Loop wakeup | `class Wakeup { int pollfd(); void signal(); void drain(); }` | `eventfd` is Linux-only |
+| Peer identity | `Result<PeerId> peer_of(int fd)` | `SO_PEERCRED` vs `LOCAL_PEERCRED` |
+| Single instance | `Result<Lock> acquire(string_view runtime_dir)` | abstract sockets are Linux-only |
+| Runtime dir | `string runtime_dir()`, `string data_dir()` | XDG vs `~/Library` |
+
+## API mapping
+
+| Linux (today) | macOS | Windows (later) | Note |
+|---|---|---|---|
+| `posix_fallocate(fd,0,n)` | `fcntl(F_PREALLOCATE)` then `ftruncate(n)` | `SetFileValidData` / `SetEndOfFile` | macOS needs the `ftruncate`; `F_PREALLOCATE` alone does not set size. Fall back to `ftruncate` on failure, exactly as the Linux path already does for `EOPNOTSUPP` |
+| `posix_fadvise(DONTNEED)` | no equivalent — **no-op** | `FILE_FLAG_NO_BUFFERING` | Do **not** substitute `F_NOCACHE`: it changes caching for the whole descriptor, not a written range. A no-op is honest; record it |
+| `fdatasync(fd)` | `fcntl(fd, F_FULLFSYNC)`, fall back to `fsync` | `FlushFileBuffers` | `fsync` on macOS does **not** guarantee the drive flushed. `.veloxpart` resume integrity depends on this — use `F_FULLFSYNC` |
+| `pwrite` | same | `WriteFile` + `OVERLAPPED` | POSIX, no work |
+| `O_NOFOLLOW` | same | `FILE_FLAG_OPEN_REPARSE_POINT` | POSIX, no work |
+| `eventfd` | self-pipe (`pipe2`/`O_NONBLOCK|O_CLOEXEC`) | event object | `poll(2)` already used, so a pipe read-end drops straight in |
+| `timerfd` | `poll()` timeout computed from the next deadline | waitable timer | Simplest port: the loop already has a deadline set |
+| `SO_PEERCRED` + `struct ucred` | `getpeereid(fd,&uid,&gid)` | named-pipe token | Same-UID check is the security property; keep it |
+| abstract socket `\0velox-daemon-<hash>` | socket file in the runtime dir + `flock(LOCK_EX|LOCK_NB)` | named mutex | macOS has no abstract namespace. Must unlink stale sockets on start — the abstract version got that free |
+| `$XDG_RUNTIME_DIR` | `$TMPDIR` (per-user, already private) | `%LOCALAPPDATA%` | macOS has no XDG runtime dir |
+| `$XDG_DATA_HOME` | `~/Library/Application Support/Velox` | `%APPDATA%` | |
+| `libsecret` / Secret Service | Keychain (`Security.framework`) | DPAPI / Credential Manager | Behind the credential-store seam. CLAUDE.md §4 still applies: never SQLite, never logs |
+| `systemd` user units | `launchd` plist (`~/Library/LaunchAgents`) | Service/Task Scheduler | Phase 3 |
+
+## Build dependencies
+
+| Ubuntu | macOS (Homebrew) |
+|---|---|
+| `qt6-base-dev`, `qt6-svg-dev`, `qt6-tools-dev` | `qt@6` |
+| `libcurl4-openssl-dev` | system libcurl, or `curl` |
+| `libsqlite3-dev` | system sqlite, or `sqlite` |
+| `libssl-dev` | `openssl@3` (set `OPENSSL_ROOT_DIR`) |
+| `libsecret-1-dev` | **none** — Keychain is in the SDK |
+| `nlohmann-json3-dev` | `nlohmann-json` |
+| `nodejs`, `npm` | `node` |
+
+The root `CMakeLists.txt` currently does `pkg_check_modules(LIBSECRET REQUIRED ...)`
+unconditionally. That must become Linux-only, or macOS cannot configure at all. It is the
+single hard blocker for a first macOS build.
+
+## What does not change
+
+`contracts/` and both generated clients, `tools/mockd`, `tests/conformance`,
+`tools/testserver`, the whole extension, and every hostile-mode expectation. If a port
+tempts you to change a fixture or a schema, stop — that is a contract change and it goes
+through PROTO.
+
+## Verification gates
+
+- **Phase 0 done:** Ubuntu suite still 57/57, and `git diff` shows only moves behind seams.
+- **Phase 1 done:** `veloxd`, `velox`, `libveloxcore` build on macOS; core unit tests pass.
+- **Phase 2 done:** conformance (mockd **and** live veloxd) and the engine hostile-mode
+  matrix pass on macOS; one real download completes with a matching SHA-256.
+- **Phase 3 done:** `.dmg` or Homebrew formula installs and runs on a clean machine.

docs/adr/0020-cross-platform-strategy.md

@@ -0,0 +1,72 @@
+# ADR 0020 — Porting to macOS (and later Windows) without forking the tree
+
+Status: **accepted** · Supersedes nothing · Applies to every lane
+
+## Context
+
+The project was built Ubuntu-first and says so in its own project description. A survey of
+`main` (2026-09-15, 57/57 green) found the Linux-specific surface is far smaller than the
+Ubuntu-first framing suggests — roughly ten files, all of them already isolated at the
+bottom of the stack:
+
+| Area | Linux-only thing | Files |
+|---|---|---|
+| `core/io` | `posix_fallocate`, `posix_fadvise`, `fdatasync` | `sparse_file.cpp` |
+| `core/meta` | `fdatasync` | `veloxpart.cpp` |
+| `daemon/rpc` | `eventfd` | `event_loop.{hpp,cpp}`, `main.cpp` |
+| `daemon/rpc` | `timerfd` | `main.cpp`, `sched/scheduler.hpp` |
+| `daemon/rpc` | `SO_PEERCRED` / `struct ucred` | `uds_server.{hpp,cpp}` |
+| `daemon/rpc` | abstract-namespace socket lock | `single_instance.{hpp,cpp}` |
+| `daemon/rpc` | `$XDG_RUNTIME_DIR` layout | `runtime_dir.{hpp,cpp}` |
+| build | `libsecret-1` is `REQUIRED` at the root | `CMakeLists.txt` |
+
+Everything else is already portable: the event loop is `poll(2)` (POSIX, not `epoll`), the
+transfer engine is libcurl, the GUI is Qt 6, `tools/mockd` and `tests/conformance` are
+Node, `tools/testserver` is stdlib Python, and the extension is a WebExtension.
+
+## Decision
+
+**One tree, one logic, per-OS backends behind seams.**
+
+1. **No `#ifdef` in logic.** Platform differences live behind a narrow interface, with one
+   implementation file per OS. A reader of `sparse_file.cpp` must not need to know which OS
+   they are on. `#ifdef` is allowed only inside a `platform/<os>/` file.
+2. **Linux is the reference implementation.** A port may never change Linux behaviour. The
+   gate is mechanical: the full suite stays green on Ubuntu, and the Linux backend keeps
+   the same syscalls it uses today. If a port needs a semantic change, that is an ADR of
+   its own, not a port commit.
+3. **Seams are introduced by the owning lane, on Linux, before any port work.** CORE
+   introduces `core`'s seams; DAEMON introduces `daemon`'s. This is a pure refactor: move
+   the existing Linux code behind the interface unchanged, prove the suite is still green.
+4. **A new lane, PORT, owns only the per-OS backends** — `**/platform/<os>/**`. It never
+   writes logic, and it does not own `cmake/`: the central OS detection module is root
+   build infrastructure and stays with PKG/QA, who must land it during Phase 0 because
+   Phase 0's own seams consume it. PORT owning it would deadlock — PORT cannot start until
+   Phase 0 lands, and Phase 0 cannot select platform sources without it. That keeps CLAUDE.md §1 intact: CORE
+   still owns `core/`'s logic, DAEMON still owns `daemon/`'s, and macOS work can proceed in
+   parallel without cross-lane writes.
+5. **`VELOX_OS_*` is set once, centrally**, in `cmake/platform.cmake` (owned by PKG/QA),
+   and platform sources are selected there — not by globbing, not per-target ad hoc.
+6. **Optional dependencies are gated, never removed.** `libsecret` stays `REQUIRED` on
+   Linux and is replaced by Keychain on macOS behind the same credential-store seam.
+
+## Phases
+
+- **Phase 0 — on Ubuntu, by CORE + DAEMON.** Introduce the seams and move today's Linux
+  code behind them. Zero behaviour change; 57/57 stays green. No macOS code exists yet.
+- **Phase 1 — on macOS, by PORT.** Implement the macOS backends until `veloxd`, `velox`
+  and `libveloxcore` build and their unit tests pass.
+- **Phase 2 — parity.** `tools/testserver`, conformance and the engine's hostile-mode
+  matrix pass on macOS. The GUI builds against Qt 6 for macOS.
+- **Phase 3 — packaging.** `.dmg`/Homebrew, `launchd` instead of `systemd`, and the macOS
+  native-messaging manifest locations. Deliberately last.
+
+## Consequences
+
+- Windows later is the same shape: a third backend directory, no new strategy. The seams
+  chosen here are POSIX-flavoured but interface-level, so a Win32 backend is additive.
+- CI must build both, or macOS rots silently. Until a macOS runner exists, Phase 1 is
+  verified by hand on the porting machine and the Linux gate stays authoritative.
+- The cost is one indirection at the bottom of the I/O and RPC stacks. Measured against
+  the alternative — `#ifdef` drift, or a forked repo that diverges in a month — it is
+  cheap.

docs/agents/AGENT-DAEMON.md

@@ -43,7 +43,11 @@ Read `contracts/`, `core/include/`, `docs/`. Never write in `core/`, `gui/`, or
    output. Build this early: it is how you test the daemon before the GUI exists.
 9. **`velox-nmhost`** — 4-byte-length-prefixed stdio ⇄ Unix socket pump. **Under 300 lines,
    zero business logic**, and it must exit cleanly when Firefox closes the pipe. Install
-   manifests to all four locations listed in `docs/05` §4.
+   manifests to the three real locations in `docs/05` §4 / `docs/adr/0003` — not four:
+   spike S1 found `~/snap/firefox/common/.mozilla/native-messaging-hosts/` (the intuitive
+   "inside the snap" path) is not actually read by snap Firefox, and corrected `docs/05` §4
+   down from its original four-location list. `packaging/nativehost/README.md` has the
+   current table.
 
 ## Definition of done (M1)
 - Passes the full conformance suite as a server, over **both** transports.

docs/agents/AGENT-PORT.md

@@ -0,0 +1,51 @@
+# Agent brief — PORT (per-OS backends)
+
+**Starts after Phase 0: CORE and DAEMON have landed the seams on Ubuntu.**
+
+## You own
+```
+core/src/**/platform/<os>/**        daemon/src/**/platform/<os>/**
+packaging/macos/**                  packaging/windows/**
+```
+You may read everything. **You never write logic** — not in `core/src/io/*.cpp`, not in
+`daemon/src/rpc/*.cpp`, not in `gui/`, never in `contracts/`.
+
+## Read first
+`docs/adr/0020-cross-platform-strategy.md`, then `docs/08-porting.md` — it has the seam
+list, the API mapping and the dependency table. Then `CLAUDE.md`.
+
+## The one rule that matters
+**Linux behaviour never changes.** If your port needs a seam that doesn't exist, or a seam
+whose shape is wrong, you do not widen it yourself — you file a request with the owning
+lane (CORE for `core/`, DAEMON for `daemon/`) exactly as every other lane does. A port that
+quietly edits shared logic is how one tree becomes two.
+
+## Build order (macOS)
+1. **Configure at all.** The root `CMakeLists.txt` requires `libsecret-1` unconditionally;
+   that must be Linux-gated before anything else compiles. PKG/QA owns that, and also owns
+   `cmake/platform.cmake` (the `VELOX_OS_*` detection and `velox_platform_sources()`); both
+   land in Phase 0, before you start. File anything else you need with them.
+2. `core/` — `preallocate`, `advise_dontneed`, `flush_durable`. Use `F_FULLFSYNC`, not
+   `fsync`: `.veloxpart` resume correctness depends on a real flush.
+3. `daemon/` — `Wakeup` (self-pipe), `peer_of` (`getpeereid`), `instance_lock` (socket file
+   + `flock`, with stale-socket cleanup), `runtime_dir` (`$TMPDIR`, `~/Library/Application
+   Support/Velox`).
+4. Unit tests green for both lanes.
+5. `tools/testserver` (stdlib Python) and conformance (Node) — expected to run unchanged.
+   If they don't, that's a bug worth reporting, not patching around.
+6. GUI against Qt 6 for macOS.
+7. Packaging — `.dmg`/Homebrew, `launchd`, native-messaging manifest locations. Last.
+
+## Definition of done (Phase 1–2)
+- macOS builds `libveloxcore`, `veloxd` and `velox` with `-Wall -Wextra -Werror`.
+- Core and daemon unit tests pass on macOS.
+- One real download completes on macOS against `tools/testserver` with a SHA-256 that
+  matches the server's reference — the same gate the Linux vertical slice passed.
+- `kill -9` mid-download, restart, resume completes and the checksum still matches.
+- The Ubuntu suite is **still** green from the same commit.
+
+## Do not
+- Do not add `#ifdef` outside a `platform/<os>/` file.
+- Do not change a fixture, a schema or a hostile-mode expectation to make macOS pass.
+- Do not substitute `F_NOCACHE` for `posix_fadvise(DONTNEED)` — see `docs/08` for why.
+- Do not start with packaging.

nmhost/CMakeLists.txt

@@ -0,0 +1,14 @@
+# nmhost/ — velox-nmhost, the Firefox native-messaging host. Owned by lane DAEMON.
+#
+# Deliberately dependency-free: no veloxd_* library, no nlohmann_json, no SQLite. It is a
+# byte-level pump between two framings (see src/main.cpp's own header comment) and runs
+# unconfined outside Firefox's sandbox (ADR 0003) — the less it links, the less there is to
+# go wrong running from wherever a snap/deb/flatpak install puts it.
+
+add_executable(velox-nmhost src/main.cpp)
+target_compile_features(velox-nmhost PRIVATE cxx_std_23)
+target_compile_options(velox-nmhost PRIVATE -Wall -Wextra -Wpedantic -Werror)
+
+if(VELOX_BUILD_TESTS AND EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/tests/CMakeLists.txt)
+    add_subdirectory(tests)
+endif()

nmhost/src/main.cpp

@@ -0,0 +1,203 @@
+// velox-nmhost — Firefox native-messaging host. A dumb pump between two framings, nothing
+// else: stdin/stdout speak Firefox's own protocol (a 4-byte native-byte-order length
+// prefix, then that many bytes of UTF-8 JSON); $XDG_RUNTIME_DIR/velox/velox.sock speaks
+// veloxd's own NDJSON (one '\n'-terminated JSON value per line, daemon/src/rpc/ndjson.hpp).
+// Reframing between the two is the entire job.
+//
+// Runs unconfined outside Firefox's snap sandbox with the real $HOME and
+// $XDG_RUNTIME_DIR (ADR 0003 §Q2) — see packaging/nativehost/ for the manifest this is
+// installed as, and which native-messaging-hosts directory actually gets read by which
+// Firefox flavour.
+//
+// No business logic: no JSON parsing (frames are pure byte spans; only the length prefix
+// and the line boundary matter here), no retry/backoff (the extension re-launches a fresh
+// host on its own reconnect), no protocol version check (veloxd and the extension settle
+// that between themselves once the pump hands their bytes through). Exits the moment
+// either side closes.
+
+#include <fcntl.h>
+#include <sys/socket.h>
+#include <sys/un.h>
+#include <unistd.h>
+
+#include <cerrno>
+#include <cstdint>
+#include <cstdlib>
+#include <cstring>
+#include <poll.h>
+#include <string>
+
+namespace {
+
+// Firefox's own cap (host -> browser) is 1 MiB; this is just a sanity backstop against a
+// runaway peer so a malformed stream can't grow a buffer without bound.
+constexpr std::size_t kMaxFrameBytes = 8 * 1024 * 1024;
+
+std::string socket_path() {
+    const char* xdg = std::getenv("XDG_RUNTIME_DIR");
+    std::string base = (xdg != nullptr && xdg[0] != '\0') ? xdg
+                                                          : ("/run/user/" + std::to_string(::getuid()));
+    if (!base.empty() && base.back() == '/') base.pop_back();
+    return base + "/velox/velox.sock";
+}
+
+int connect_socket(const std::string& path) {
+    const int fd = ::socket(AF_UNIX, SOCK_STREAM | SOCK_CLOEXEC, 0);
+    if (fd < 0) return -1;
+    sockaddr_un addr{};
+    addr.sun_family = AF_UNIX;
+    if (path.size() + 1 > sizeof(addr.sun_path)) {
+        ::close(fd);
+        return -1;
+    }
+    std::memcpy(addr.sun_path, path.c_str(), path.size());
+    if (::connect(fd, reinterpret_cast<sockaddr*>(&addr), sizeof(addr)) != 0) {
+        ::close(fd);
+        return -1;
+    }
+    return fd;
+}
+
+// Reads everything currently available into `buf`. true on EOF, false otherwise (short
+// reads / EAGAIN just return with whatever was appended).
+bool read_available(int fd, std::string& buf) {
+    char chunk[65536];
+    for (;;) {
+        const ssize_t n = ::read(fd, chunk, sizeof(chunk));
+        if (n > 0) {
+            buf.append(chunk, static_cast<std::size_t>(n));
+            continue;
+        }
+        if (n == 0) return true;  // EOF
+        if (errno == EAGAIN || errno == EWOULDBLOCK) return false;
+        if (errno == EINTR) continue;
+        return true;  // treat any other error as if the peer hung up
+    }
+}
+
+// Writes as much of `buf` as the fd accepts right now, trimming what was sent. Returns
+// false on a hard error (peer gone); EAGAIN is not an error, just "try again later".
+bool flush_some(int fd, std::string& buf) {
+    while (!buf.empty()) {
+        const ssize_t n = ::write(fd, buf.data(), buf.size());
+        if (n > 0) {
+            buf.erase(0, static_cast<std::size_t>(n));
+            continue;
+        }
+        if (n < 0 && (errno == EAGAIN || errno == EWOULDBLOCK)) return true;
+        if (n < 0 && errno == EINTR) continue;
+        return false;
+    }
+    return true;
+}
+
+// stdin frames (4-byte length + payload) -> NDJSON lines appended to `to_socket`.
+// Malformed (oversized) length is a hard stop.
+bool drain_stdin_frames(std::string& in, std::string& to_socket) {
+    for (;;) {
+        if (in.size() < 4) return true;
+        std::uint32_t len;
+        std::memcpy(&len, in.data(), 4);
+        if (len > kMaxFrameBytes) return false;
+        if (in.size() < 4 + len) return true;
+        to_socket.append(in, 4, len);
+        to_socket.push_back('\n');
+        in.erase(0, 4 + len);
+    }
+}
+
+// NDJSON lines from the socket -> stdout frames (4-byte length + payload) appended to
+// `to_stdout`.
+bool drain_socket_lines(std::string& in, std::string& to_stdout) {
+    for (;;) {
+        const auto nl = in.find('\n');
+        if (nl == std::string::npos) {
+            if (in.size() > kMaxFrameBytes) return false;
+            return true;
+        }
+        std::string_view line(in.data(), nl);
+        if (!line.empty() && line.back() == '\r') line.remove_suffix(1);
+        if (line.size() > kMaxFrameBytes) return false;
+        const auto len = static_cast<std::uint32_t>(line.size());
+        to_stdout.append(reinterpret_cast<const char*>(&len), 4);
+        to_stdout.append(line);
+        in.erase(0, nl + 1);
+    }
+}
+
+}  // namespace
+
+int main() {
+    const int sock = connect_socket(socket_path());
+    if (sock < 0) return 1;  // daemon not running / socket missing: nothing to pump
+
+    // Every fd this pumps must be non-blocking: read_available()'s own drain loop keeps
+    // calling read() until it actually sees EAGAIN, and a blocking fd never returns that —
+    // it just blocks inside the "drain what's available" loop instead of going back to
+    // poll(), which stalls the whole pump the moment one side has more to send than fits
+    // in a single read().
+    ::fcntl(0, F_SETFL, ::fcntl(0, F_GETFL) | O_NONBLOCK);
+    ::fcntl(1, F_SETFL, ::fcntl(1, F_GETFL) | O_NONBLOCK);
+    ::fcntl(sock, F_SETFL, ::fcntl(sock, F_GETFL) | O_NONBLOCK);
+
+    std::string stdin_buf, stdout_buf, socket_in_buf, socket_out_buf;
+    bool stdin_eof = false;
+
+    // Three distinct descriptors, not two: stdin (0) and stdout (1) are separate pipes
+    // (read-only and write-only respectively — never the same fd, even though they sit
+    // next to each other in a shell's mental model of "the process's stdio"), plus the
+    // bidirectional socket.
+    enum { kStdin, kStdout, kSock };
+    for (;;) {
+        pollfd fds[3] = {
+            {0, 0, 0},
+            {1, 0, 0},
+            {sock, 0, 0},
+        };
+        if (!stdin_eof) fds[kStdin].events |= POLLIN;
+        if (!stdout_buf.empty()) fds[kStdout].events |= POLLOUT;
+        fds[kSock].events |= POLLIN;
+        if (!socket_out_buf.empty()) fds[kSock].events |= POLLOUT;
+
+        // Nothing left to wait for: both directions exhausted.
+        if (fds[kStdin].events == 0 && fds[kStdout].events == 0 && fds[kSock].events == 0) break;
+
+        const int n = ::poll(fds, 3, -1);
+        if (n < 0) {
+            if (errno == EINTR) continue;
+            break;
+        }
+
+        if (fds[kStdout].revents & POLLOUT) {
+            if (!flush_some(1, stdout_buf)) break;  // Firefox closed our stdout
+        }
+        if (fds[kSock].revents & POLLOUT) {
+            if (!flush_some(sock, socket_out_buf)) break;
+        }
+        if (fds[kStdin].revents & (POLLIN | POLLHUP)) {
+            if (read_available(0, stdin_buf)) stdin_eof = true;
+            if (!drain_stdin_frames(stdin_buf, socket_out_buf)) break;
+        }
+        if (fds[kSock].revents & (POLLIN | POLLHUP)) {
+            const bool socket_eof = read_available(sock, socket_in_buf);
+            if (!drain_socket_lines(socket_in_buf, stdout_buf)) break;
+            if (socket_eof) {
+                // The daemon is gone. Flush whatever we already turned into stdout
+                // frames, then stop — there is nothing left to relay either direction.
+                (void)flush_some(1, stdout_buf);
+                break;
+            }
+        }
+        if ((fds[kStdin].revents | fds[kStdout].revents | fds[kSock].revents) &
+            (POLLERR | POLLNVAL))
+            break;
+
+        // Firefox closed the pipe: nothing more will ever arrive on stdin, and once our
+        // own outbound backlog drains there is nothing left to send it either. Stop
+        // rather than idle forever relaying replies nobody reads.
+        if (stdin_eof && socket_out_buf.empty()) break;
+    }
+
+    ::close(sock);
+    return 0;
+}

nmhost/tests/CMakeLists.txt

@@ -0,0 +1,13 @@
+# Integration test only — velox-nmhost has no internal functions worth unit-testing in
+# isolation (it's ~15 lines of byte-shuffling helpers around one poll() loop); what matters
+# is the real binary's observable behaviour over real pipes and a real socket.
+
+add_executable(velox_nmhost_pump_test pump_test.cpp)
+target_compile_features(velox_nmhost_pump_test PRIVATE cxx_std_23)
+target_compile_options(velox_nmhost_pump_test PRIVATE -Wall -Wextra -Wpedantic -Werror)
+target_compile_definitions(velox_nmhost_pump_test PRIVATE
+    VELOX_NMHOST_BIN="$<TARGET_FILE:velox-nmhost>")
+add_dependencies(velox_nmhost_pump_test velox-nmhost)
+
+add_test(NAME nmhost.pump COMMAND velox_nmhost_pump_test)
+set_tests_properties(nmhost.pump PROPERTIES TIMEOUT 30)

nmhost/tests/check.hpp

@@ -0,0 +1,54 @@
+#pragma once
+
+// Minimal test harness: CHECK accumulates failures, TEST_MAIN reports and sets the exit
+// code. Copied from daemon/tests/check.hpp rather than shared across a build-dependency —
+// nmhost is deliberately dependency-free, tests included.
+
+#include <cstdio>
+#include <string>
+#include <vector>
+
+namespace veloxd_test {
+
+inline std::vector<std::string>& failures() {
+    static std::vector<std::string> f;
+    return f;
+}
+inline int& checks() {
+    static int n = 0;
+    return n;
+}
+
+}  // namespace veloxd_test
+
+#define CHECK(cond)                                                                     \
+    do {                                                                               \
+        ++::veloxd_test::checks();                                                     \
+        if (!(cond)) {                                                                 \
+            ::veloxd_test::failures().push_back(std::string(__FILE__) + ":" +          \
+                                               std::to_string(__LINE__) + ": " + #cond); \
+        }                                                                              \
+    } while (0)
+
+#define CHECK_EQ(a, b)                                                                  \
+    do {                                                                               \
+        ++::veloxd_test::checks();                                                     \
+        auto _va = (a);                                                                \
+        auto _vb = (b);                                                                \
+        if (!(_va == _vb)) {                                                           \
+            ::veloxd_test::failures().push_back(std::string(__FILE__) + ":" +          \
+                                               std::to_string(__LINE__) + ": " + #a +  \
+                                               " == " + #b);                           \
+        }                                                                              \
+    } while (0)
+
+#define TEST_MAIN()                                                                     \
+    int main() {                                                                        \
+        run();                                                                          \
+        for (const auto& f : ::veloxd_test::failures()) std::printf("FAIL %s\n", f.c_str()); \
+        std::printf("%d/%d checks passed\n",                                            \
+                    ::veloxd_test::checks() -                                           \
+                        static_cast<int>(::veloxd_test::failures().size()),             \
+                    ::veloxd_test::checks());                                           \
+        return ::veloxd_test::failures().empty() ? 0 : 1;                               \
+    }

nmhost/tests/pump_test.cpp

@@ -0,0 +1,148 @@
+// Integration test for velox-nmhost: spawns the real binary, feeds it a framed stdin
+// message, answers over a fake Unix socket standing in for veloxd, and checks what comes
+// back out on stdout — plus that it exits promptly once stdin closes.
+
+#include <sys/socket.h>
+#include <sys/stat.h>
+#include <sys/un.h>
+#include <sys/wait.h>
+#include <unistd.h>
+
+#include <cstdint>
+#include <cstdlib>
+#include <cstring>
+#include <ctime>
+#include <string>
+
+#include "check.hpp"
+
+namespace {
+
+std::string make_temp_dir() {
+    char tmpl[] = "/tmp/velox-nmhost-test-XXXXXX";
+    const char* dir = ::mkdtemp(tmpl);
+    return dir ? dir : "/tmp";
+}
+
+// Firefox's own framing: 4-byte native-byte-order length, then that many bytes.
+std::string frame(const std::string& payload) {
+    std::uint32_t len = static_cast<std::uint32_t>(payload.size());
+    std::string out(reinterpret_cast<char*>(&len), 4);
+    out += payload;
+    return out;
+}
+
+// Reads exactly one framed message off `fd`, blocking. Empty on EOF/short read.
+std::string read_frame(int fd) {
+    std::uint32_t len = 0;
+    std::size_t got = 0;
+    while (got < 4) {
+        const ssize_t n = ::read(fd, reinterpret_cast<char*>(&len) + got, 4 - got);
+        if (n <= 0) return {};
+        got += static_cast<std::size_t>(n);
+    }
+    std::string payload(len, '\0');
+    got = 0;
+    while (got < len) {
+        const ssize_t n = ::read(fd, payload.data() + got, len - got);
+        if (n <= 0) return {};
+        got += static_cast<std::size_t>(n);
+    }
+    return payload;
+}
+
+bool write_all(int fd, const std::string& s) {
+    std::size_t off = 0;
+    while (off < s.size()) {
+        const ssize_t n = ::write(fd, s.data() + off, s.size() - off);
+        if (n <= 0) return false;
+        off += static_cast<std::size_t>(n);
+    }
+    return true;
+}
+
+}  // namespace
+
+void run() {
+    const std::string dir = make_temp_dir();
+    const std::string velox_dir = dir + "/velox";
+    CHECK(::mkdir(velox_dir.c_str(), 0700) == 0);
+    const std::string sock_path = velox_dir + "/velox.sock";
+
+    // A bare listening socket standing in for veloxd.
+    const int listen_fd = ::socket(AF_UNIX, SOCK_STREAM, 0);
+    CHECK(listen_fd >= 0);
+    sockaddr_un addr{};
+    addr.sun_family = AF_UNIX;
+    std::memcpy(addr.sun_path, sock_path.c_str(), sock_path.size());
+    CHECK(::bind(listen_fd, reinterpret_cast<sockaddr*>(&addr), sizeof(addr)) == 0);
+    CHECK(::listen(listen_fd, 1) == 0);
+
+    int child_stdin[2];   // [0] read (child), [1] write (parent)
+    int child_stdout[2];  // [0] read (parent), [1] write (child)
+    CHECK(::pipe(child_stdin) == 0);
+    CHECK(::pipe(child_stdout) == 0);
+
+    ::setenv("XDG_RUNTIME_DIR", dir.c_str(), 1);
+
+    const pid_t pid = ::fork();
+    CHECK(pid >= 0);
+    if (pid == 0) {
+        ::dup2(child_stdin[0], 0);
+        ::dup2(child_stdout[1], 1);
+        ::close(child_stdin[0]);
+        ::close(child_stdin[1]);
+        ::close(child_stdout[0]);
+        ::close(child_stdout[1]);
+        ::close(listen_fd);
+        ::execl(VELOX_NMHOST_BIN, "velox-nmhost", nullptr);
+        ::_exit(127);
+    }
+    ::close(child_stdin[0]);
+    ::close(child_stdout[1]);
+
+    // Accept nmhost's connection.
+    const int conn = ::accept(listen_fd, nullptr, nullptr);
+    CHECK(conn >= 0);
+
+    // stdin (framed) -> socket (NDJSON line).
+    CHECK(write_all(child_stdin[1], frame(R"({"hello":1})")));
+    char line[256] = {};
+    ssize_t n = ::read(conn, line, sizeof(line) - 1);
+    CHECK(n > 0);
+    CHECK_EQ(std::string(line, static_cast<std::size_t>(n)), std::string("{\"hello\":1}\n"));
+
+    // socket (NDJSON line) -> stdout (framed).
+    CHECK(write_all(conn, "{\"world\":2}\n"));
+    const std::string got = read_frame(child_stdout[0]);
+    CHECK_EQ(got, std::string(R"({"world":2})"));
+
+    // A second round trip on the same connection, to prove buffering across calls works
+    // (not just "the first message happens to line up with one read()").
+    CHECK(write_all(child_stdin[1], frame(R"({"again":3})")));
+    n = ::read(conn, line, sizeof(line) - 1);
+    CHECK(n > 0);
+    CHECK_EQ(std::string(line, static_cast<std::size_t>(n)), std::string("{\"again\":3}\n"));
+
+    // Firefox closes the pipe: nmhost must exit promptly rather than hang.
+    ::close(child_stdin[1]);
+    int status = 0;
+    pid_t waited = -1;
+    for (int i = 0; i < 50 && waited != pid; ++i) {
+        waited = ::waitpid(pid, &status, WNOHANG);
+        if (waited == pid) break;
+        struct timespec ts{0, 20'000'000};  // 20ms
+        ::nanosleep(&ts, nullptr);
+    }
+    CHECK_EQ(waited, pid);
+    if (waited == pid) CHECK(WIFEXITED(status));
+
+    ::close(conn);
+    ::close(child_stdout[0]);
+    ::close(listen_fd);
+    ::unlink(sock_path.c_str());
+    ::rmdir(velox_dir.c_str());
+    ::rmdir(dir.c_str());
+}
+
+TEST_MAIN()

packaging/nativehost/README.md

@@ -0,0 +1,97 @@
+# packaging/nativehost — the Firefox native-messaging manifest
+
+Owner: lane DAEMON (`packaging/nativehost/**` is explicitly DAEMON's per `CLAUDE.md`'s
+lane table, unlike the rest of `packaging/`, which is PKG/QA's). This directory holds the
+manifest `velox-nmhost` (built from `nmhost/`) is registered under, and this file is the
+one place that says exactly which on-disk locations need it and why — the
+"install manifests to all four locations" line in `docs/agents/AGENT-DAEMON.md`'s build
+order predates `docs/adr/0003-native-messaging-under-snap.md`, which cut that down to
+three *real* ones. Read the ADR before touching this list; it is the empirical spike, not
+a guess.
+
+## The manifest
+
+`com.velox.host.json` in this directory is the template:
+
+```json
+{
+  "name": "com.velox.host",
+  "description": "Velox download manager — native messaging bridge to veloxd",
+  "path": "/usr/libexec/velox/velox-nmhost",
+  "type": "stdio",
+  "allowed_extensions": ["velox@velox.download"]
+}
+```
+
+- `name` is what the extension calls `browser.runtime.connectNative("com.velox.host")`
+  with (`extension/src/background/transport/native.ts`'s `HOST_NAME`) — do not rename one
+  without the other.
+- `allowed_extensions` must match `extension/manifest.json`'s
+  `browser_specific_settings.gecko.id` exactly (`[email protected]`). A mismatch here is
+  a silent failure: Firefox reports "no such native application" as if the manifest didn't
+  exist at all, which reads exactly like a missing-file bug and is easy to chase in the
+  wrong place.
+- `path` is `/usr/libexec/velox/velox-nmhost` — the `.deb`'s install layout
+  (`docs/07-packaging.md`). **Every other packaging format needs a different `path`** —
+  the binary doesn't live at that absolute path inside a Flatpak sandbox or an AppImage
+  mount, so whatever installs the manifest for those formats must rewrite this field to
+  wherever it actually put the binary, not ship this file verbatim. That substitution is
+  the packaging step's job, not this template's.
+
+## Where it has to be written, and why
+
+Per ADR 0003 (spike S1, run against the real snap Firefox on the target machine — not a
+guess about how confinement "should" work):
+
+| Location | Covers | Why |
+|---|---|---|
+| `~/.mozilla/native-messaging-hosts/com.velox.host.json` | **deb/tarball Firefox, and snap Firefox** | The one location snap Firefox 154 actually reads. Firefox's snap launches native-messaging hosts *outside* the sandbox with the user's real `$HOME` — so this is not "the deb location that happens to also work"; it is *the* snap location, full stop. `~/snap/firefox/common/.mozilla/native-messaging-hosts/` (the intuitive "inside the snap" path) is **not** read by Firefox 154 snap — confirmed empirically, not inferred. |
+| `/usr/lib/mozilla/native-messaging-hosts/com.velox.host.json` | **deb/tarball Firefox only** | System-wide, so it covers every user on the box for a real (non-snap) Firefox install — but confirmed **not** read by snap Firefox. Do not treat this path as "covers snap too"; that was the wrong assumption `docs/05` §4 corrected in the same commit as the ADR. |
+| `~/.var/app/org.mozilla.firefox/.mozilla/native-messaging-hosts/com.velox.host.json` | **Flatpak Firefox** | Flatpak's own sandboxed home. Untested on this machine (Firefox here is the snap, not flatpak) — carried over from `docs/05` §4's original four-location list, which this table otherwise supersedes. Verify before relying on it in a release checklist. |
+
+That is three real locations, not four — the fourth
+(`~/snap/firefox/common/.mozilla/native-messaging-hosts/`) was the pre-ADR guess the spike
+disproved. If `docs/agents/AGENT-DAEMON.md`'s build order still says "four locations" when
+you read this, it is stale; this table is the current source of truth alongside the ADR
+itself.
+
+## What this means for `postinst` (PKG/QA's file, not this one)
+
+This directory ships the manifest template and documents the target paths; it does not
+install anything itself (`packaging/**` outside this one directory is PKG/QA's — see
+`CLAUDE.md`'s lane table — and `postinst` specifically is a `.deb`-packaging concern this
+lane doesn't own the file for). For whoever writes it:
+
+- **The two `~/`-relative locations are per-user.** `postinst` runs as root, once, at
+  install time — it does not run once per user session. It needs either a real-user
+  enumeration at install time (every UID with a home directory and no login shell of
+  `/usr/sbin/nologin`-style exclusions, roughly what `deluser --remove-home` scripts already
+  have to reason about) or a first-run hook that runs as the logged-in user (a systemd user
+  unit's `ExecStartPre`, or the GUI's own first-run wizard) and writes its own manifest the
+  first time it starts. `docs/adr/0003`'s own follow-up section flagged this as open; it
+  still is.
+- **`/usr/lib/mozilla/native-messaging-hosts/` is the only one of the three that's a plain
+  root-owned, install-time write** — no per-user enumeration needed for that one.
+- **Detect snap Firefox and say so.** `docs/07-packaging.md` already commits to this
+  ("detects whether Firefox is a snap... prints... a one-line note that the extension will
+  pair over loopback") — the detection matters here specifically because if Firefox turns
+  out to be neither deb/tarball nor a snap this spike covers (a genuinely unknown or future
+  packaging of Firefox), silently trusting `NativeTransport` to work is exactly the failure
+  mode ADR 0003 exists to prevent. `WebSocketTransport` is the guaranteed fallback either
+  way (ADR 0003's own decision) — nothing breaks if the manifest doesn't land correctly, it
+  just means the extension pairs over loopback instead of the (opportunistic, not
+  required) native path.
+- **`docs/07-packaging.md`'s own install layout line currently lists only the
+  system-wide `/usr/lib/mozilla/...` path.** That line is correct as far as it goes (it *is*
+  one of the three locations, and the only pure root-owned one) but reads as if it were the
+  whole story for native messaging; it predates this file and the ADR. Worth a line
+  pointing here so the two documents don't quietly disagree — PKG/QA's call, not edited
+  here since `docs/07` is PKG/QA's own file.
+
+## The binary
+
+`nmhost/` builds `velox-nmhost` — see that directory's own `src/main.cpp` for what it does
+(a byte-level pump, no protocol logic) and `docs/05-extension-spec.md` §4 for the two
+transports it sits behind. It is deliberately dependency-free (not linked against any
+`veloxd_*` library) so wherever a packaging format's sandbox puts it, it has nothing else
+to go looking for at runtime.

packaging/nativehost/com.velox.host.json

@@ -0,0 +1,7 @@
+{
+  "name": "com.velox.host",
+  "description": "Velox download manager — native messaging bridge to veloxd",
+  "path": "/usr/libexec/velox/velox-nmhost",
+  "type": "stdio",
+  "allowed_extensions": ["[email protected]"]
+}

packaging/systemd/README.md

@@ -0,0 +1,65 @@
+# packaging/systemd — the user unit pair for veloxd
+
+Provided by lane DAEMON (`daemon/docs/AGENT-DAEMON.md` build step 7: "systemd user units
+— `velox.service` + `velox.socket` for socket activation"), for PKG/QA to install per
+`docs/07-packaging.md`'s layout:
+
+```
+/usr/lib/systemd/user/velox.service
+/usr/lib/systemd/user/velox.socket
+```
+
+`packaging/` outside `nativehost/` is PKG/QA's per `CLAUDE.md`'s lane table; these two
+files are here because they are inputs to that packaging step, not a claim on the rest of
+the directory — same relationship `packaging/nativehost/` already has.
+
+## Why both files, and what `RuntimeDirectory=` is doing
+
+`velox.socket` binds `$XDG_RUNTIME_DIR/velox/velox.sock` **before `veloxd` ever runs** and
+hands the daemon the already-listening fd at startup (`daemon/src/rpc/systemd_activation.cpp`
+implements the receiving half — `LISTEN_PID`/`LISTEN_FDS`, fd 3 — without a `libsystemd`
+link). Two things this buys over the daemon binding its own socket on every start:
+
+- **No window where a client gets `ECONNREFUSED`.** The socket exists and queues
+  connections from the moment `velox.socket` is active, not from whenever `veloxd`
+  finishes starting up — this is the actual point of socket activation, not just "start
+  on demand."
+- **Cold-boot ordering is free.** Nothing has to wait for `veloxd` to be ready before the
+  GUI, the CLI, or a native-messaging host can attempt a connection; the kernel queues it.
+
+`RuntimeDirectory=velox` on the socket unit creates `%t/velox` (mode 0700) before the
+`ListenStream=` bind — without it, binding fails outright the first time (nothing has
+created the parent directory yet). `veloxd` itself creates that same directory
+(`ensure_private_dir` in `runtime_dir.cpp`) for the case where it's started directly,
+outside systemd (`./veloxd` in a terminal, still supported and how most of this daemon's
+own testing runs) — the two paths converge on the same directory with the same mode
+either way.
+
+`UdsServer::start()` (`daemon/src/rpc/uds_server.cpp`) checks for the activated fd first
+and, if present, skips create/bind/chmod/listen entirely — the socket file's lifecycle then
+belongs to the unit (including `RemoveOnStop=yes` on stop), not to the daemon. Falls back
+to binding its own socket exactly as before when not socket-activated (a manual run, or a
+distro that ships the daemon without the unit files).
+
+## What was deliberately left out
+
+`velox.service` does **not** set `ProtectSystem=`, `ProtectHome=`, or `ReadWritePaths=`.
+`saveTo.allowedRoots` is user-configurable to anywhere on the filesystem — an external
+drive, a second mount, anywhere `fs/safepath.hpp`'s own canonicalize-and-check accepts —
+not a fixed set of directories a unit file could enumerate ahead of time. A filesystem-level
+sandbox here would turn a legitimately-configured save location into an opaque
+`EROFS`/`EACCES` the daemon can't explain, in place of its own clear `-32011` — worse than
+no sandbox, specifically for a download manager. `NoNewPrivileges=yes` is kept: it has no
+such trade-off.
+
+## Verifying socket activation without a real install
+
+`systemd-analyze verify --user velox.service velox.socket` checks unit-file syntax (it
+will complain that `/usr/bin/veloxd` and the `velox(1)` man page don't exist on a dev
+box that hasn't installed the package — expected, not a unit bug). To exercise the actual
+activation handshake without installing anything: bind a Unix socket, `dup2` it onto fd 3,
+fork, set `LISTEN_PID=<child pid>` and `LISTEN_FDS=1` in the child's environment, clear
+`FD_CLOEXEC` on fd 3, and `execve` `veloxd` — a real `session.hello` round-trips over that
+fd with no `bind()`/`listen()` call ever happening inside the daemon for that run. This is
+exactly what `velox.socket`'s `Requires=`/`ExecStart` sequence does in production; systemd
+supplies the fd, `veloxd` doesn't know the difference.

packaging/systemd/velox.service

@@ -0,0 +1,38 @@
+[Unit]
+Description=Velox download manager daemon
+Documentation=man:velox(1)
+# Socket activation (velox.socket) means this unit does not need to be enabled or started
+# directly for the RPC transport to come up on demand — the first connection attempt after
+# boot starts veloxd with the listening socket already bound (see velox.socket's own
+# comment). Requires=/After= still matter for a manual `systemctl --user start velox`.
+Requires=velox.socket
+After=velox.socket
+
+# Never more than one real instance for this user regardless of how it was started — the
+# abstract-socket single-instance lock (main.cpp, keyed off the resolved runtime dir) is
+# the actual enforcement; this just keeps systemd itself from racing two starts.
+StartLimitIntervalSec=60
+StartLimitBurst=5
+
+[Service]
+Type=simple
+ExecStart=/usr/bin/veloxd
+# main.cpp's SIGTERM handler stops the event loop and falls through to a clean shutdown
+# (flushes buffers, closes the store, releases the single-instance lock) — the default
+# KillSignal=SIGTERM and TimeoutStopSec are already the right shape for that; no
+# ExecStop/KillMode override needed.
+Restart=on-failure
+RestartSec=2
+
+# Hardening deliberately stops here, not at ProtectSystem=/ProtectHome=/ReadWritePaths=:
+# saveTo.allowedRoots is user-configurable to anywhere (an external drive, a second
+# mount — fs/safepath.hpp is the daemon's own validation boundary, not a fixed set of
+# directories a unit file could enumerate up front). A filesystem-level sandbox here would
+# silently turn a legitimately-configured save location into an opaque EROFS/EACCES the
+# daemon can't explain, instead of its own clear -32011 — worse than no sandbox, for a
+# download manager specifically. NoNewPrivileges is free of that trade-off.
+NoNewPrivileges=yes
+
+[Install]
+WantedBy=default.target
+Also=velox.socket

packaging/systemd/velox.socket

@@ -0,0 +1,28 @@
+[Unit]
+Description=Velox download manager — RPC socket
+
+[Socket]
+# %t is $XDG_RUNTIME_DIR for a user unit — the exact path veloxd itself resolves
+# (daemon/src/rpc/runtime_dir.cpp's resolve_runtime_dir), and the exact path velox(1)
+# resolves too (default_socket_path() in cli/src/client.cpp). All three must agree; this
+# is the one line that has to.
+ListenStream=%t/velox/velox.sock
+
+# RuntimeDirectory creates %t/velox (mode 0700, this user's own) before binding, so the
+# ListenStream= path above always has somewhere to land — veloxd itself does the same
+# thing (ensure_private_dir) when it creates the directory unassisted (the non-activated
+# path, e.g. a manual `veloxd` run outside systemd).
+RuntimeDirectory=velox
+RuntimeDirectoryMode=0700
+
+# Same-UID-only, matching the socket's own authorization once a connection is accepted
+# (SO_PEERCRED, checked again in uds_server.cpp regardless of this mode bit — belt and
+# braces, not a substitute for it).
+SocketMode=0600
+
+# The socket file belongs to systemd's socket-activation state, not to whatever's on disk
+# from a previous boot; remove it on stop so a stale entry never shadows the next start.
+RemoveOnStop=yes
+
+[Install]
+WantedBy=sockets.target