diff --git a/docs/daemon-startup-hardening.md b/docs/daemon-startup-hardening.md
new file mode 100644
index 0000000..5aa315d
--- /dev/null
+++ b/docs/daemon-startup-hardening.md
@@ -0,0 +1,495 @@
+# Daemon Startup Hardening — Implementation Plan
+
+Eight verified edge-case defects in how ObsidianDragon brings up (and watches) the
+`dragonxd` daemon at launch. Each entry is a buildable fix: the defect (with exact
+line references), the chosen approach, the call sites, a representative change, and how
+to verify it.
+
+- **Scope:** full-node startup path (`--lite` excludes the embedded daemon entirely).
+- **Source:** line references are exact against branch `dev` @ `45b652f`.
+- **Provenance:** findings verified by direct source read; each fix designed by an
+  independent agent grounded in the cited files, with a sequencing pass for ordering,
+  shared helpers, and merge conflicts.
+
+**Severity:** 2 High, 6 Medium · **Effort:** ≈ 25–35 engineering-hours · **7 landing steps.**
+
+Status legend: ☐ not started · ◐ in progress · ☑ landed & verified
+
+---
+
+## Recommended rollout sequence
+
+A real dependency order, not a checklist. The daemon-lifecycle cluster lands first
+because it makes the `State::Error` / `crash_count_` contract trustworthy — which the
+connect-stall panel and the lock gate both build on. The filesystem cluster lands
+around a single shared helper. The connectivity-breaking security flip lands last.
+
+| Step | Finding(s) | Site | Why here | Status |
+|------|-----------|------|----------|--------|
+| 1 | **F1** | `embedded_daemon.cpp` · `isRunning()` | Smallest/highest-severity; establishes the reliable Error/crash-count transition steps 3 & 6 depend on. | ☑ |
+| 2 | **F2** | `embedded_daemon.cpp` · `startProcess()` | Same file family, different function; test the F1+F2 pair together with `kill -SEGV` / bad-binary repros. | ☑ |
+| 3 | **F4** | `embedded_daemon.cpp` · `start()` | After F1/F2 so crash-count semantics are settled; its bail deliberately stays out of the crash path. | ☑ |
+| 4 | **F7** | `util/platform` · `connection.cpp` | Structural owner of the fs-error idiom + `ConnectionConfig` that F5/F6/F8 reuse. | ☐ |
+| 5 | **F6 + F5** | `app.cpp` · `verifySaplingParams()` | Same `startEmbeddedDaemon` / `verifySaplingParams` block; land together. | ☐ |
+| 6 | **F3** | `app.cpp` · `renderLoadingOverlay()` | After F1 — panel is guarded off during `State::Error` (owned by the crash-count hint). | ☐ |
+| 7 | **F8** | `connection.cpp` · `tryConnect()` | Largest; only connectivity-breaking default flip — land last, with release notes. | ☐ |
+
+---
+
+## F1 — Double-`waitpid` race can swallow a daemon crash
+
+**Severity:** High · **Effort:** S (~1–2h) · **Status:** ☑ landed & verified
+
+### The defect
+`EmbeddedDaemon::isRunning()` (`embedded_daemon.cpp:1136`, POSIX branch) calls
+`waitpid(WNOHANG)` — from the **UI thread, nearly every frame** — racing
+`monitorProcess()`'s own reap at `:1244`. `waitpid` is one-shot: if the UI thread wins,
+the monitor never decodes the exit, so `crash_count_` never increments, `State::Error`
+never fires, and the 3-strike auto-restart cap (`app_network.cpp:479`) is defeated. The
+sibling `XmrigManager::isRunning()` (`xmrig_manager.cpp:512`) already fixed exactly this
+with an atomic read.
+
+### The fix
+Make `isRunning()` read the existing `std::atomic<State> state_` (member at
+`embedded_daemon.h:253`) instead of calling `waitpid`, leaving `monitorProcess()` as the
+sole reaper. Predicate is `Running || Stopping` — `Stopping` must stay "alive" because
+`stop()`'s graceful/SIGTERM wait loops poll `isRunning()` before the process has exited.
+
+### Files touched
+- `src/daemon/embedded_daemon.cpp` — `isRunning()`, POSIX branch (~1136)
+
+### Core change
+```cpp
+bool EmbeddedDaemon::isRunning() const   // POSIX branch
+{
+    // Read the atomic state_ instead of waitpid() — monitorProcess() is the
+    // sole reaper. Previously both threads reaped; if the UI thread won, the
+    // monitor never saw the exit (crash_count_ / exit code / Error all lost).
+    if (process_pid_ <= 0) return false;
+
+    State s = state_.load(std::memory_order_relaxed);
+    // Stopping stays "alive": stop()'s wait loops poll isRunning() while
+    // state_ == Stopping, before the process has actually terminated.
+    return (s == State::Running || s == State::Stopping);
+}
+```
+
+### Verification
+- Manual: `kill -SEGV` the daemon 10–20×; the monitor must report the exit and increment `crash_count_` every time (previously intermittent).
+- Regression: a normal Settings-driven stop still escalates SIGTERM→SIGKILL (the `Stopping` predicate).
+- Not unit-testable (real fork/exec/waitpid) — consistent with the no-process-spawn harness.
+
+### Dependencies
+Mirrors `XmrigManager::isRunning()`. Flags a separate latent hazard (out of scope):
+`stop()`'s final blocking `waitpid` (`:1220`) can still race a mid-sleep monitor
+iteration — file as its own ticket.
+
+---
+
+## F2 — exec-after-fork silent failure: "Running" for a daemon that never started
+
+**Severity:** High · **Effort:** S (~2–3h) · **Status:** ☑ landed & verified
+
+### The defect
+In `startProcess()` (`embedded_daemon.cpp:957–1061`, POSIX) the parent runs
+`process_pid_ = pid; return true;` **unconditionally** after `fork()` — with no
+exec-status handshake. On a non-executable / wrong-arch / corrupt binary the child's
+`execv` fails and it `_exit(127)`s, but `start()` has already set `State::Running`
+(`:565`). The real cause never reaches `last_error_`; it surfaces later, generically,
+as "exited unexpectedly (exit code 127)".
+
+### The fix
+Add a **close-on-exec self-pipe** handshake — `pipe() + fcntl(FD_CLOEXEC)`, deliberately
+**not** `pipe2()` (macOS lacks it; the POSIX branch is shared). The child writes `errno`
+only on `execv` failure; a successful exec closes the write end for free. Parent reads:
+EOF ⇒ success; 4 bytes ⇒ reap the zombie, set a precise `last_error_` ("not executable
+or wrong architecture"), and return `false` so `start()` never reports Running. EINTR-safe
+on both ends. Also comments the unchecked parent-side `setpgid` at `:1053`.
+
+### Files touched
+- `src/daemon/embedded_daemon.cpp` — `startProcess()` parent read path
+- `src/daemon/embedded_daemon.cpp` — child `execv`-failure write (~1043)
+- `src/daemon/embedded_daemon.cpp` — `setpgid` best-effort comment (~1053)
+
+### Core change
+```cpp
+// Self-pipe exec handshake (pipe()+FD_CLOEXEC; NOT pipe2 — macOS lacks it).
+int execpipe[2]; pipe(execpipe);
+fcntl(execpipe[0], F_SETFD, FD_CLOEXEC);
+fcntl(execpipe[1], F_SETFD, FD_CLOEXEC);
+
+pid_t pid = fork();
+if (pid == 0) {                        // child
+    close(execpipe[0]);
+    /* setpgid / chdir / dup2 / argv … */
+    execv(binary_path.c_str(), argv.data());
+    int e = errno;                     // execv failed
+    while (write(execpipe[1], &e, sizeof e) < 0 && errno == EINTR) {}
+    _exit(127);
+}
+
+close(execpipe[1]);                    // parent: must close or read() never EOFs
+int child_errno = 0, total = 0;
+for (;;) {                             // EOF ⇒ exec ok; 4 bytes ⇒ exec failed
+    ssize_t n = read(execpipe[0], (char*)&child_errno + total, sizeof(int) - total);
+    if (n == 0) break;
+    if (n < 0) { if (errno == EINTR) continue; break; }
+    if ((total += n) >= (int)sizeof(int)) break;
+}
+close(execpipe[0]);
+if (total >= (int)sizeof(int)) {       // exec never happened
+    waitpid(pid, nullptr, 0);          // reap the zombie
+    last_error_ = "dragonxd could not be executed: " +
+                  std::string(strerror(child_errno)) +
+                  " — not executable or wrong architecture";
+    return false;                      // start() no longer reports Running
+}
+```
+
+### Verification
+- Point at a `chmod -x` / wrong-arch file → `start()` returns false immediately, precise message, no leftover zombie.
+- Success path: real binary still starts with no perceptible added latency.
+- Optional pure `formatExecFailureError(errno)` helper for a `test_phase4.cpp` unit test.
+
+### Dependencies
+F1 (same function family; sequence F1→F2). **Highest-risk mistake:** forgetting
+`FD_CLOEXEC` makes every successful start hang the parent read forever.
+
+---
+
+## F4 — Stale datadir-lock start → restart storm that wedges the UI
+
+**Severity:** Medium · **Effort:** S (~3–5h) · **Status:** ☑ landed & verified
+
+### The defect
+`start()` (`embedded_daemon.cpp:466`) gates only on the RPC port (`:482`), never on
+`isDaemonProcessRunning()` (`:1292`). A graceful shutdown frees the port but keeps the
+datadir `.lock` for up to ~90s. A rapid stop→start spawns a daemon that dies "Cannot
+obtain a lock on data directory" — routed to the generic crash path. With a ~4s retry
+cadence, **three lock races in ~12s exhaust the 3-strike budget** and wedge the UI long
+before the lock actually clears.
+
+### The fix
+Fail-fast with a **short bounded local wait (~300ms), not a 90s block**. After the port
+bail, consult `isDaemonProcessRunning()` — gated by `!skip_port_check_` and exempt when
+`override_datadir_` is set, so the isolated migrate-to-seed daemon still works. A pure
+`evaluateDatadirLockGate()` returns a **distinct non-crash Error** that never increments
+`crash_count_`. The connect loop's own retry then absorbs the transient.
+
+### Files touched
+- `src/daemon/embedded_daemon.h` — decision struct, helper decl, poll constants
+- `src/daemon/embedded_daemon.cpp` — `start()` gate + `evaluateDatadirLockGate()`
+
+### Core change
+```cpp
+static StartLockGateDecision evaluateDatadirLockGate(
+        bool skipPortCheck, bool isolatedOverride, bool stillRunningAfterWait) {
+    if (skipPortCheck || isolatedOverride) return {true,  ""};   // migrate-to-seed exempt
+    if (!stillRunningAfterWait)            return {true,  ""};
+    return {false, "A previous dragonxd is still shutting down and holding the "
+                   "data directory lock. Retrying shortly…"};
+}
+
+// start() — after the isPortInUse() bail, before setState(Starting):
+if (!skip_port_check_ && override_datadir_.empty()) {
+    bool stillLocked = false;                       // ~300ms bounded wait, NOT ~90s
+    for (int i = 0; i < kDatadirLockWaitMaxPolls; ++i) {
+        if (!isDaemonProcessRunning()) { stillLocked = false; break; }
+        stillLocked = true;
+        std::this_thread::sleep_for(std::chrono::milliseconds(kDatadirLockWaitPollMs));
+    }
+    auto gate = evaluateDatadirLockGate(false, false, stillLocked);
+    if (!gate.proceed) { setState(State::Error, gate.errorMessage); return false; }
+}
+```
+
+### Verification
+- Unit: `evaluateDatadirLockGate()` across the skip / isolated / still-running matrix.
+- Manual: rapid restart into a lingering lock → distinct message, no crash-cap wedge.
+- Migrate-to-seed second daemon still starts (isolated exemption).
+
+### Dependencies
+F1/F2 (must not touch `crash_count_`; wording must not collide with the monitor's
+"exited unexpectedly"). Same TU, different function.
+
+---
+
+## F5 — Extraction / copy write-failures never surfaced up front
+
+**Severity:** Medium · **Effort:** S (~2–3h) · **Status:** ☐
+
+### The defect
+`startEmbeddedDaemon()` discards `extractEmbeddedResources()`'s `bool` return
+(`app.cpp:4152`) and the second copy-fallback loop drops `copy_file`'s `error_code`
+entirely (`:4236`). Only Sapling params **existence** is re-checked — never the daemon
+binary/CLI/tx/asmap. A disk-full or truncated `dragonxd` write falls straight through to
+spawn and fails opaquely. The innermost write already returns `false`
+(`embedded_resources.cpp:307`) — the signal is simply thrown away.
+
+### The fix
+Minimal, surgical wiring — no new abstraction. Capture the extraction return and, on
+failure, set `daemon_status_ = TR("sb_daemon_extract_failed")` and `return false` before
+spawning. In the second copy loop, check `ec` after each `copy_file`, track `copyFailed`,
+and abort with a dir-parameterized `sb_daemon_files_failed`. An **absent source** stays
+fine (optional files); only an actual `error_code` counts. Written so F6/F7 slot in later
+without re-touching this control flow.
+
+### Files touched
+- `src/app.cpp` — `startEmbeddedDaemon()` extraction check (~4152)
+- `src/app.cpp` — second copy-fallback loop (~4210–4242)
+- `src/util/i18n.cpp` + `res/lang/*.json` — 2 additive keys
+
+### Core change
+```cpp
+// stop discarding the extraction result (~4152)
+if (!resources::extractEmbeddedResources()) {
+    daemon_status_ = TR("sb_daemon_extract_failed");   // disk full / permission denied
+    return false;                                      // abort before spawning
+}
+
+// second copy-fallback loop — was dropping ec entirely (~4236)
+bool copyFailed = false;
+for (const char* name : { "asmap.dat", "dragonxd", "dragonx-cli", "dragonx-tx" }) {
+    fs::path dst = fs::path(daemon_dir) / name;
+    if (fs::exists(dst)) continue;                      // already present — skip
+    for (const auto& dir : searchDirs) {
+        fs::path src = fs::path(dir) / name;
+        if (!fs::exists(src)) continue;                 // absent source is OK, not a failure
+        fs::copy_file(src, dst, ec);
+        if (ec) { copyFailed = true; ec.clear(); }
+        break;
+    }
+}
+if (copyFailed) {
+    char buf[512];
+    snprintf(buf, sizeof buf, TR("sb_daemon_files_failed"), daemon_dir.c_str());
+    daemon_status_ = buf;
+    return false;                                       // don't fall through to spawn
+}
+```
+
+### Verification
+- Unit: `extractEmbeddedResources()` returns false without embedded resources.
+- Extract the copy loop into a testable helper; force one dst write to fail (dst is an existing directory).
+- Manual: near-full tmpfs / read-only dir → clear status, daemon controller never constructed.
+
+### Dependencies
+Shares the `daemon_status_` surfacing convention with F6; its early-return pattern is the
+template F7 matches. Open item: remove truncated dst files so a retry re-copies.
+
+---
+
+## F6 — Sapling params validated by existence/size only, never hashed
+
+**Severity:** Medium · **Effort:** S (~3–5h) · **Status:** ☐
+
+### The defect
+`verifySaplingParams()` (`connection.cpp:123`) only calls `fs::exists()`;
+`resourceNeedsUpdate()` (`embedded_resources.cpp:250`) is size-only. On Linux (no
+embedded resources) a **truncated-but-present** param passes and is handed to the daemon,
+which then fails to build shielded proofs mid-operation — far from the real cause.
+
+### The fix
+Add a pinned `{ filename → size, sha256 }` table (one source of truth, cross-referenced
+to `scripts/build-lite-backend-artifact.sh`) and hash-check each param after the
+existence check, reusing the existing `util::sha256Hex` (no second implementation). Since
+these are ~48 MB, **cache the result** via a `.sapling_verified` marker keyed on
+`size:mtime` — re-hash only when the stat line changes, so startup isn't slowed.
+
+### Files touched
+- `src/rpc/connection.h` — `verifySaplingParams` decl
+- `src/rpc/connection.cpp` — digest table, marker helpers, rewrite
+
+### Core change
+```cpp
+// connection.cpp — pinned known-good digests
+// (source of truth: scripts/build-lite-backend-artifact.sh ensure_sapling_params)
+constexpr SaplingParamDigest kSaplingParamDigests[] = {
+  { "sapling-spend.params",  47958396, "8e48ffd2…efc13" },
+  { "sapling-output.params",  3592860, "2f0ebbcb…fb0e4" },
+};
+
+bool Connection::verifySaplingParams() {
+    // existence check (unchanged) …
+    // cache: skip re-hashing a ~48 MB file unless size:mtime changed
+    if (readMarkerMatches(marker, statLines)) return true;
+    for (auto& d : kSaplingParamDigests)
+        if (util::sha256Hex(bytes) != d.sha256) return false;   // reuse existing helper
+    writeMarker(marker, statLines);
+    return true;
+}
+```
+
+### Verification
+- Unit: good params pass; truncated / wrong-bytes rejected; marker cache short-circuits re-hash unless size/mtime changed. Real temp-file fixtures (matches existing `sha256Hex` tests).
+
+### Dependencies
+F7 (reuse fs-error idiom; shares the `startEmbeddedDaemon`/`verifySaplingParams` block).
+Third caller of the existing `util::sha256Hex`.
+
+---
+
+## F7 — Directory-create errors universally ignored on the daemon-env path
+
+**Severity:** Medium · **Effort:** S (~3–4h) · **Status:** ☐
+
+### The defect
+Five startup directory-create sites either drop the `error_code` or use the throwing
+overload with no `catch`: `main.cpp:730`, `connection.cpp:216` (can throw **uncaught**
+through its callers), `embedded_resources.cpp:270`, `app.cpp:4172`/`4218`. A read-only
+home or permission-denied yields a confusing "conf missing" / "binary not found"
+downstream — or an uncaught `filesystem_error` — instead of a clear cause.
+
+### The fix
+One shared, non-throwing `Platform::ensureDirectory(dir, outError)` in
+`util/platform.{h,cpp}` that produces a single consistent message. Replace all five
+sites; `autoDetectConfig()` moves off the throwing overload and sets a new
+`ConnectionConfig::dir_error` that its four callers check and bail on. This is the
+**structural owner** of the fs-error idiom that F5 and F6 reuse.
+
+### Files touched
+- `src/util/platform.h` / `.cpp` — `ensureDirectory()`
+- `src/rpc/connection.h` / `.cpp` — `dir_error` + `autoDetectConfig`
+- `main.cpp`, `app.cpp`, `app_network.cpp`, `app_wizard.cpp`, `settings_page.cpp`, `embedded_resources.cpp` — 5 sites + 4 callers
+- `tests/test_phase4.cpp` — `TestPlatformEnsureDirectory`
+
+### Core change
+```cpp
+// util/platform.cpp — one shared, non-throwing helper
+bool Platform::ensureDirectory(const std::string& dir, std::string* outError) {
+    std::error_code ec;
+    if (std::filesystem::is_directory(dir, ec)) return true;
+    ec.clear();
+    std::filesystem::create_directories(dir, ec);
+    if (ec) {
+        if (outError)
+            *outError = "Cannot create " + dir + ": " + ec.message() +
+                        ". Check permissions / free space.";
+        return false;
+    }
+    return true;
+}
+// Replaces 5 ad-hoc sites; autoDetectConfig() now sets ConnectionConfig::dir_error,
+// and its 4 callers bail on it.
+```
+
+### Verification
+- Unit `TestPlatformEnsureDirectory`: existing dir → true; fresh nested → created; POSIX unwritable → false + message.
+- All four `autoDetectConfig` callers tolerate `dir_error`. Pre-App-init site (main.cpp) reports via stderr / MessageBox.
+
+### Dependencies
+**Owns** `Platform::ensureDirectory` (used by F5, F6) and the `ConnectionConfig`
+extension (coordinated with F8). Land before F5/F6/F8.
+
+---
+
+## F8 — Plaintext-remote RPC credential transmission is warn-only
+
+**Severity:** Medium · **Effort:** M (~6–9h) · **Status:** ☐
+
+### The defect
+A remote `rpchost` without `rpctls=1` sends Basic-auth `rpcuser:rpcpassword` over
+cleartext HTTP. `tryConnect()` (`app_network.cpp:314`) only shows a **dismissible
+warning** then proceeds — a local-network MITM sees the credentials. Compounding it,
+`isLocalHost()`'s naive `rfind("127.",0)==0` misclassifies `127.evil.com` as local,
+suppressing even the warning.
+
+### The fix
+Change the policy to **refuse-by-default with an explicit, persisted opt-in** — a
+`rpcallowplaintext=1` conf key (for hand-editors) and a Settings toggle. Block the
+connect and show a **blocking modal** explaining the risk and how to enable TLS or opt
+in; localhost is unaffected. Tighten `isLocalHost()` to exact `127.x.y.z` / `::1` /
+`localhost` via `isExactIPv4Loopback()`. **Back-compat:** default off ⇒ existing remote
+users hit a hard stop until they opt in — **ship with prominent release notes.**
+
+### Files touched
+- `src/rpc/connection.h` / `.cpp` — `isLocalHost`, `allow_plaintext_remote`, `parseConfFile`
+- `src/config/settings.h` / `.cpp` — persisted opt-in
+- `src/app_network.cpp`, `src/app.h` — refuse + modal dispatch
+- `src/ui/windows/plaintext_remote_rpc_dialog.h` — new blocking modal
+- `src/ui/pages/settings_page.cpp` — toggle UI
+
+### Core change
+```cpp
+// Tightened loopback test — "127.evil.com" is NOT local
+bool Connection::isLocalHost(const std::string& host) {
+    std::string h = stripBrackets(lowercase(host));
+    return h == "localhost" || h == "::1" || isExactIPv4Loopback(h);  // exact 127.x.y.z
+}
+
+// Refuse-by-default with an explicit, persisted opt-in
+const bool plaintextRemote  = rpc::Connection::usesPlaintextRemote(config);
+const bool plaintextAllowed = config.allow_plaintext_remote          // rpcallowplaintext=1
+                           || settings_.getAllowPlaintextRemoteRpc(); // Settings toggle
+if (plaintextRemote && !plaintextAllowed) {
+    connection_status_ = TR("sb_plaintext_remote_blocked");
+    showPlaintextRemoteRpcDialog(config.host + ":" + config.port);    // blocking modal
+    return;                                                           // no creds sent
+}
+```
+
+### Verification
+- Unit: `isLocalHost` — `127.evil.com` false, `127.0.0.1`/`::1`/`localhost` true; `allowsPlaintextRemote` honors conf key + settings flag.
+- Manual: remote plaintext blocked; modal fires; opt-in persists across restart.
+
+### Dependencies
+F7 (second extender of `ConnectionConfig`/`parseConfFile`; land after so the struct grows
+once). Wire `renderPlaintextRemoteRpcDialog` into the app modal-dispatch list.
+
+---
+
+## Shared helpers & coordination points
+
+| Helper | Purpose | Used by |
+|--------|---------|---------|
+| `Platform::ensureDirectory()` | Single non-throwing directory-create with one consistent message; replaces five ad-hoc sites. Owned by F7. | F7, F5, F6 |
+| `ConnectionConfig` extension | Coordination point, not a function: F7 adds `dir_error`, F8 adds `allow_plaintext_remote`. Land F7→F8 so it grows once per step. | F7, F8 |
+| `util::sha256Hex` *(existing)* | Already-compiled, curl-free SHA-256. F6 becomes its third caller — no second hash routine. | F6 |
+| `connectHasStalled()` *(new, pure)* | Stall predicate split out of the ImGui/App code for unit testing, per the `*_updater_core.cpp` precedent. | F3 |
+| `evaluateDatadirLockGate()` *(new, pure)* | Lock-gate decision as `{proceed, message}` from three booleans — unit-testable without real process/fs I/O. | F4 |
+
+## F3 — Unbounded connect spinner (deferred to step 6)
+
+**Severity:** Medium · **Effort:** S (~3–5h) · **Status:** ☐
+
+The connect loop retries forever while `!state_.connected` (`app.cpp:1239`);
+`loading_timer_` only animates the spinner. Stamp `connect_stall_since_` when
+"reachable but not ready" is first seen; a pure `connectHasStalled()` helper (new
+`util/connect_stall.h`, default 45s from `ui.toml`) flips `state_.connect_stalled` at
+threshold, and `renderLoadingOverlay()` shows a "Taking longer than expected" panel with
+Retry / Restart daemon / Open console (full-node gated). The background retry keeps
+firing — recovery clears the panel automatically. Guarded off while the daemon is in
+`State::Error` (owned by F1's crash-count hint). Full detail lives in the sequencing/
+design record; see the shared-helper table above.
+
+---
+
+## Cross-cutting notes
+
+- **One TU, three functions.** `embedded_daemon.cpp` is edited by F1 (`isRunning`),
+  F2 (`startProcess`) and F4 (`start`) — no literal hunk overlap, but land in order to
+  keep "monitorProcess is the sole reaper" coherent.
+- **Connection struct grows twice.** `connection.h/.cpp` is touched by F6, F7 and F8;
+  F7 and F8 both extend `ConnectionConfig` and `parseConfFile` — highest collision risk.
+  Sequence F7→F6→F8.
+- **Testability split.** The three new pure predicates all get `tests/test_phase4.cpp`
+  coverage. F1/F2's fork/exec/waitpid changes are **not** unit-testable — they rely on
+  manual `kill` / non-executable-binary repros, consistent with the no-process-spawn harness.
+- **i18n is additive-only.** Add each finding's English keys to `strings_`, then run
+  `scripts/add_missing_translations.py` **once at the very end**
+  (`json.dump indent=4, sort_keys=True, ensure_ascii=False`) — never bulk-regenerate a
+  `res/lang/*.json`.
+- **F8 is a breaking default flip.** Refuse-plaintext-by-default stops existing
+  remote-RPC users cold until they opt in. Lands last, gated behind a persisted opt-in,
+  with release notes calling out the new `rpcallowplaintext` key and the Settings toggle.
+- **Latent hazard, out of scope.** F1 surfaces (but doesn't fix) a second
+  double-`waitpid` window between `stop()`'s final blocking reap (`:1220`) and a
+  mid-sleep monitor iteration — file it as its own ticket.
+
+---
+
+## Progress log
+
+- **F1** — ☑ landed: `isRunning()` (POSIX) now reads the atomic `state_` (predicate `Running || Stopping`) instead of calling `waitpid`, leaving `monitorProcess()` the sole reaper. Clean build (all targets link); `ctest` 1/1 passing. Not unit-testable — needs the manual `kill -SEGV` repro before release.
+- **F2** — ☑ landed: `startProcess()` (POSIX) now creates a `FD_CLOEXEC` self-pipe before `fork()`; the child writes `errno` to it on `execv` failure, the parent reads EOF-vs-errno and, on failure, reaps the zombie + sets a precise `last_error_` ("not executable or wrong architecture") + returns `false` (so `start()` no longer reports `Running` for a daemon that never started). Parent-side `setpgid` is now best-effort with a `DEBUG_LOGF` on failure. Clean build; `ctest` 1/1 passing. Not unit-testable — needs the manual non-executable / wrong-arch-binary repro before release.
+- **F4** — ☑ landed: `start()` now gates on a lingering datadir lock after the port bail. When `!skip_port_check_ && override_datadir_.empty()`, it polls `isDaemonProcessRunning()` with a bounded ~300 ms wait (3 × 100 ms, breaks early), then a pure header-inline `evaluateDatadirLockGate()` decides: if a sibling `dragonxd` is still alive it bails with a distinct **non-crash** `State::Error` ("…holding the data directory lock. Retrying shortly…") that never touches `crash_count_`, so the 3-strike cap can't trip; the connect loop's retry resumes once the lock clears. Isolated migrate-to-seed starts are exempt. New `testDatadirLockGate` unit test (5 assertions, proceed/bail/2× exempt) added to `test_phase4.cpp`. Clean build; `ctest` 1/1 passing.
diff --git a/src/daemon/embedded_daemon.cpp b/src/daemon/embedded_daemon.cpp
index 2e67fd6..2add767 100644
--- a/src/daemon/embedded_daemon.cpp
+++ b/src/daemon/embedded_daemon.cpp
@@ -488,6 +488,34 @@ bool EmbeddedDaemon::start(const std::string& binary_path)
         return false;
     }
     external_daemon_detected_ = false;
+
+    // A previous dragonxd can release the RPC port well before it releases the datadir
+    // .lock — a graceful shutdown can take up to ~90s (see isDaemonProcessRunning). Starting
+    // into a still-held lock spawns a process that dies instantly with "Cannot obtain a lock
+    // on data directory"; the crash monitor reports that generically and, three times in
+    // ~12s, that is enough to trip the 3-strike restart cap before the lock's ~90s life
+    // elapses. Gate on the process actually still being alive, with a SHORT bounded wait
+    // (not the full ~90s — start() runs on the UI thread). Isolated starts (migrate-to-seed:
+    // skip_port_check_ / -datadir override) are exempt; they run their own datadir+port.
+    {
+        constexpr int kDatadirLockWaitPollMs   = 100;
+        constexpr int kDatadirLockWaitMaxPolls = 3;   // ~300ms total, breaks early on exit
+        bool stillRunning = false;
+        if (!skip_port_check_ && override_datadir_.empty()) {
+            stillRunning = isDaemonProcessRunning();
+            for (int i = 0; stillRunning && i < kDatadirLockWaitMaxPolls; ++i) {
+                std::this_thread::sleep_for(std::chrono::milliseconds(kDatadirLockWaitPollMs));
+                stillRunning = isDaemonProcessRunning();
+            }
+        }
+        const StartLockGateDecision gate =
+            evaluateDatadirLockGate(skip_port_check_, !override_datadir_.empty(), stillRunning);
+        if (!gate.proceed) {
+            VERBOSE_LOGF("[INFO] %s\n", gate.errorMessage);
+            setState(State::Error, gate.errorMessage);
+            return false;
+        }
+    }
     
     setState(State::Starting, "Looking for dragonxd binary...");
     
@@ -962,18 +990,38 @@ bool EmbeddedDaemon::startProcess(const std::string& binary_path, const std::vec
         last_error_ = "Failed to create pipe: " + std::string(strerror(errno));
         return false;
     }
+
+    // Self-pipe used purely as an exec-success/failure handshake, separate from
+    // the stdout pipe above. Both ends are close-on-exec, so a successful execv()
+    // closes the write end for free (parent reads EOF); on execv() failure the
+    // child writes errno here, so the parent learns synchronously instead of
+    // reporting State::Running for a child that never became dragonxd. We use
+    // pipe()+FD_CLOEXEC (not pipe2) because this POSIX branch is shared with
+    // macOS, which has no pipe2().
+    int execpipe[2];
+    if (pipe(execpipe) == -1) {
+        last_error_ = "Failed to create exec-status pipe: " + std::string(strerror(errno));
+        close(pipefd[0]);
+        close(pipefd[1]);
+        return false;
+    }
+    fcntl(execpipe[0], F_SETFD, FD_CLOEXEC);
+    fcntl(execpipe[1], F_SETFD, FD_CLOEXEC);
     
     pid_t pid = fork();
     if (pid == -1) {
         last_error_ = "Fork failed: " + std::string(strerror(errno));
         close(pipefd[0]);
         close(pipefd[1]);
+        close(execpipe[0]);
+        close(execpipe[1]);
         return false;
     }
     
     if (pid == 0) {
         // Child process
-        close(pipefd[0]);  // Close read end
+        close(pipefd[0]);   // Close read end of the stdout pipe
+        close(execpipe[0]); // Child only writes the exec-status pipe
         
         // Put child in its own process group so we can kill the entire
         // group later (including dragonxd spawned by a wrapper script).
@@ -1040,22 +1088,61 @@ bool EmbeddedDaemon::startProcess(const std::string& binary_path, const std::vec
             execv(binary_path.c_str(), argv.data());
         }
         
-        // If we get here, exec failed
-        fprintf(stderr, "execv failed: %s\n", strerror(errno));
+        // If we get here, execv() failed — the child never became dragonxd.
+        // Capture errno before fprintf/strerror can clobber it, report it to
+        // the parent over the exec-status pipe (EINTR-safe), then exit.
+        int exec_errno = errno;
+        fprintf(stderr, "execv failed: %s\n", strerror(exec_errno));
+        ssize_t w;
+        do {
+            w = write(execpipe[1], &exec_errno, sizeof(exec_errno));
+        } while (w < 0 && errno == EINTR);
         _exit(127);
     }
     
     // Parent process
-    close(pipefd[1]);  // Close write end
+    close(pipefd[1]);   // Close our copy of the stdout write end
+    close(execpipe[1]); // Must close our copy, or the read() below never sees EOF
+
+    // Exec-status handshake: EOF => execv() succeeded (its write end was closed
+    // on exec); a full sizeof(int) => execv() failed and the child sent errno.
+    int child_errno = 0;
+    size_t got = 0;
+    char* ep = reinterpret_cast<char*>(&child_errno);
+    for (;;) {
+        ssize_t n = read(execpipe[0], ep + got, sizeof(child_errno) - got);
+        if (n == 0) break;                                   // EOF: exec succeeded
+        if (n < 0) { if (errno == EINTR) continue; break; }  // other error: assume success
+        got += static_cast<size_t>(n);
+        if (got >= sizeof(child_errno)) break;               // full errno: exec failed
+    }
+    close(execpipe[0]);
+
+    if (got >= sizeof(child_errno)) {
+        // execv() never replaced the child; it fprintf'd and _exit(127)'d. Reap
+        // the already-dead zombie here — monitorProcess() is only started after
+        // this function returns true, so there is no competing reaper.
+        close(pipefd[0]);
+        int status;
+        waitpid(pid, &status, 0);
+        last_error_ = "dragonxd could not be executed: " + std::string(strerror(child_errno)) +
+                      " — not executable or wrong architecture";
+        return false;
+    }
+
     stdout_fd_ = pipefd[0];
-    
-    // Also set process group from parent side (race with child's setpgid)
-    setpgid(pid, pid);
-    
+
+    // Best-effort: the child already calls setpgid(0, 0); this parent-side call
+    // just closes the fork/exec race window. A failure here is not fatal to
+    // startup, so we log rather than abort.
+    if (setpgid(pid, pid) != 0) {
+        DEBUG_LOGF("[WARN] setpgid(%d) from parent failed: %s\n", (int)pid, strerror(errno));
+    }
+
     // Set non-blocking
     int flags = fcntl(stdout_fd_, F_GETFL, 0);
     fcntl(stdout_fd_, F_SETFL, flags | O_NONBLOCK);
-    
+
     process_pid_ = pid;
     return true;
 }
@@ -1135,17 +1222,21 @@ double EmbeddedDaemon::getMemoryUsageMB() const
 
 bool EmbeddedDaemon::isRunning() const
 {
+    // Read the atomic state_ instead of calling waitpid() here. monitorProcess()
+    // is the sole thread allowed to waitpid() process_pid_ during normal operation.
+    // Calling waitpid() from this method too (as it used to, and this is invoked
+    // from the UI thread nearly every frame) meant whichever thread reaped the
+    // child's exit first consumed the status; if isRunning() won that race,
+    // monitorProcess() never saw the exit, so crash_count_ / the decoded exit
+    // code / the State::Error transition were all silently lost. Mirrors the
+    // fix already in XmrigManager::isRunning().
     if (process_pid_ <= 0) return false;
-    
-    int status;
-    pid_t result = waitpid(process_pid_, &status, WNOHANG);
-    
-    if (result == 0) {
-        // Still running
-        return true;
-    }
-    
-    return false;
+
+    const State s = state_.load(std::memory_order_relaxed);
+    // State::Stopping is included: stop()'s graceful/SIGTERM wait loops poll
+    // isRunning() while state_ == Stopping — before the process has actually
+    // terminated — and must keep seeing "alive" to wait/escalate correctly.
+    return (s == State::Running || s == State::Stopping);
 }
 
 void EmbeddedDaemon::drainOutput()
diff --git a/src/daemon/embedded_daemon.h b/src/daemon/embedded_daemon.h
index 4021cab..bce7513 100644
--- a/src/daemon/embedded_daemon.h
+++ b/src/daemon/embedded_daemon.h
@@ -235,6 +235,32 @@ public:
      */
     static bool isDaemonProcessRunning();
 
+    /** Decision returned by evaluateDatadirLockGate(): whether start() may spawn now. */
+    struct StartLockGateDecision {
+        bool proceed = true;             // false => bail before spawning
+        const char* errorMessage = "";   // set (a string literal) when proceed == false
+    };
+
+    /**
+     * @brief Pure decision for start(): bail because a previous dragonxd still holds the
+     *        shared datadir lock? Isolated instances (skip_port_check_ / an active -datadir
+     *        override) are exempt — they run their own throwaway datadir+port and can coexist
+     *        with the main daemon. Does no process/fs I/O itself (the caller does the probing),
+     *        so it is directly unit-testable; defined inline so tests need only this header.
+     */
+    static StartLockGateDecision evaluateDatadirLockGate(bool skipPortCheck,
+                                                         bool isolatedOverride,
+                                                         bool stillRunningAfterWait)
+    {
+        if (skipPortCheck || isolatedOverride) return {true, ""};
+        if (stillRunningAfterWait) {
+            return {false,
+                    "A previous dragonxd is still shutting down and holding the data "
+                    "directory lock. Retrying shortly…"};
+        }
+        return {true, ""};
+    }
+
     /** @brief Is an arbitrary TCP port currently in use on localhost? (used to pick a free port) */
     static bool tcpPortInUse(int port);
 
diff --git a/tests/test_phase4.cpp b/tests/test_phase4.cpp
index 14cf7ed..d868e22 100644
--- a/tests/test_phase4.cpp
+++ b/tests/test_phase4.cpp
@@ -3,6 +3,7 @@
 #include "chat/chat_service.h"
 #include "chat/chat_database.h"
 #include "daemon/daemon_controller.h"
+#include "daemon/embedded_daemon.h"
 #include "data/transaction_history_cache.h"
 #include "data/address_book.h"
 #include "data/wallet_index.h"
@@ -2477,6 +2478,28 @@ void testDaemonShutdownPolicy()
     EXPECT_TRUE(bootstrap.disconnectRpc);
 }
 
+void testDatadirLockGate()
+{
+    using dragonx::daemon::EmbeddedDaemon;
+
+    // Normal start, no lingering daemon after the bounded wait → proceed.
+    auto clear = EmbeddedDaemon::evaluateDatadirLockGate(false, false, false);
+    EXPECT_TRUE(clear.proceed);
+
+    // A previous dragonxd still alive after the wait → bail with a distinct, non-crash msg.
+    auto locked = EmbeddedDaemon::evaluateDatadirLockGate(false, false, true);
+    EXPECT_TRUE(!locked.proceed);
+    EXPECT_TRUE(std::string(locked.errorMessage).find("data directory lock") != std::string::npos);
+
+    // Isolated instance via skip_port_check_ is exempt even if a sibling dragonxd is running.
+    auto skipPort = EmbeddedDaemon::evaluateDatadirLockGate(true, false, true);
+    EXPECT_TRUE(skipPort.proceed);
+
+    // Isolated instance via -datadir override is exempt even if a sibling is running.
+    auto isolated = EmbeddedDaemon::evaluateDatadirLockGate(false, true, true);
+    EXPECT_TRUE(isolated.proceed);
+}
+
 void testDaemonLifecycleExecution()
 {
     using dragonx::daemon::DaemonController;
@@ -6619,6 +6642,7 @@ int main()
     testWalletSecurityWorkflow();
     testWalletSecurityWorkflowExecutor();
     testDaemonShutdownPolicy();
+    testDatadirLockGate();
     testDaemonLifecycleExecution();
     testDaemonLifecycleAdapters();
     testConsoleTextLayout();