PARITY.md

@@ -160,8 +160,8 @@ being the one that leaks.
 |---|:--:|:--:|---|
 | Back off on repeated app-lock attempts | ✅ | ✅ | ios#41 — 3 free attempts, then 30s doubling to a 5-minute ceiling |
 | Handle a wrong clock | ✅ | ✅ | ios#45 — skew learned free from the `Date` header; expiry takes the LATER of device and relay |
-| Zero key material after use | ✅ | ⚪ | ios#39 |
-| Pin the relay's identity | ✅ | ⚪ | ios#19 |
+| Zero key material after use | ✅ | 🟡 | ios#39 — done, but the JVM guarantee is **weaker than iOS**; see below |
+| Pin the relay's identity | ✅ | ✅ | ios#19 — Android already pinned the key and drew the **wrong conclusion** from a change; see below |
 
 ⚠️ **The back-off is not a defence of the data.** It slows a person holding an
 unlocked phone. The vault and the ratchet keys are protected by the Android
@@ -185,6 +185,33 @@ the UI thread, so the two fields are `@Volatile` and updated together under a
 lock. A torn read of "skew from now, timestamp from an hour ago" would age out an
 observation just made, or trust one that had not been.
 
+🔴 **The relay pin was not a missing feature on Android — it was a wrong
+conclusion.** `Trust.verifyHeadSignature` already pinned `issuerPublicKey` the
+first time a signed log head arrived. When the key later changed, the signature
+stopped verifying and the only conclusion available at that point was
+`KeyLogForkException` — so a rotation somebody performed on purpose was reported
+to the user as *the relay showing two different histories*.
+
+Those need different responses. A fork means nothing the relay says can be
+trusted. A rotation means somebody ran a command, or somebody else is answering
+on that address, and only a person can tell which. `RelayIdentityChangedException`
+is deliberately **not** a `KeyLogForkException`, and `forkDetected` is not set —
+an alarm people cannot act on is one they learn to dismiss, which would cost the
+real fork detection its credibility too.
+
+🔴 **Zeroing is where Android cannot match iOS, and the row says 🟡 for that
+reason rather than being rounded up to ✅.** iOS has `memset_s`, specified not to
+be optimised away, and iOS does not page application memory to disk. Neither
+holds here: a compacting collector can relocate a `ByteArray` and leave the
+bytes at the old address untouched, Android has had compressed swap on by
+default for years, and `Arrays.fill` is merely "not eliminated in practice"
+rather than guaranteed. There is no way to pin a Java array, so this is a
+ceiling on what any implementation here can offer — not a gap to close later.
+
+The claim Android supports is therefore narrower and is written that way in
+`SecureBytes`: **a derived key is not left sitting in the heap after use.** Not
+"the key is gone".
+
 🔴 One divergence from iOS, and it is a bug the Swift version cannot have. iOS
 computes `30 * pow(2, steps)` in a `Double`, which saturates to infinity and
 still caps correctly. Kotlin shifting a `Long` does not: `30L shl 60` wraps

client/src/main/kotlin/io/flamenet/client/RelayIdentityPin.kt

@@ -0,0 +1,112 @@
+package io.flamenet.client
+
+/**
+ * Which relay this is, beyond which certificate it presented (ios#19).
+ *
+ * 🔴 **TLS answers "is this host who DNS says it is". It does not answer "is
+ * this the same relay I have been talking to".** A mis-issued certificate, a
+ * hijacked DNS record, or a hostile party who obtains the domain all produce a
+ * server with a perfectly valid certificate for the right name — and every one
+ * of them is a different relay wearing the same address.
+ *
+ * The relay publishes an issuer public key in its descriptor. Pinning that on
+ * first contact turns "the certificate is valid" into "this is the same relay",
+ * which is a materially stronger statement and the one that matters.
+ *
+ * 🔴 **Android already pinned this key and drew the wrong conclusion from it.**
+ * `Trust.verifyHeadSignature` pins `issuerPublicKey` the first time a signed log
+ * head arrives, and if the key later changes the signature simply fails to
+ * verify — so the client raises `KeyLogForkException`, sets `forkDetected`, and
+ * tells somebody the relay's key log does not add up.
+ *
+ * A rotated relay key is not a forked log. They need different responses: a
+ * fork means the relay is showing two different histories and nothing it says
+ * can be trusted; a rotation means somebody ran a command, or somebody else is
+ * answering on that address. Reporting the second as the first is both alarming
+ * and unactionable, and an alarm people cannot act on is one they learn to
+ * dismiss — which costs the real fork detection its credibility too.
+ *
+ * ⚠️ **A relay that publishes no key is a weaker deployment, not an attack.**
+ * Somebody self-hosting an older build has nothing to pin, and refusing to talk
+ * to them would be a worse failure than the one this prevents. That case is
+ * distinguished from a key that CHANGED, which is not weakness.
+ *
+ * ⚠️ A change is surfaced, never silently accepted and never silently refused.
+ * The same shape as a contact's identity changing: stop, say so, and let a
+ * person decide — because a legitimate rotation exists and an app that cannot
+ * survive one is an app that gets uninstalled.
+ */
+object RelayIdentityPin {
+
+    /** What a descriptor's key means against what was pinned before. */
+    sealed interface Verdict {
+        /** Nothing pinned and a key published: pin it. */
+        data class FirstSight(val offered: ByteArray) : Verdict {
+            override fun equals(other: Any?) =
+                other is FirstSight && offered.contentEquals(other.offered)
+            override fun hashCode() = offered.contentHashCode()
+        }
+
+        /** The same relay as before. */
+        data object Unchanged : Verdict
+
+        /** This deployment publishes no key. Nothing to check. */
+        data object NotPublished : Verdict
+
+        /**
+         * 🔴 A different key than the one pinned. Either a deliberate rotation
+         * or a different server answering on this address, and nothing here can
+         * tell those apart — only a person can.
+         */
+        data class Changed(val pinned: ByteArray, val offered: ByteArray) : Verdict {
+            override fun equals(other: Any?) = other is Changed &&
+                pinned.contentEquals(other.pinned) && offered.contentEquals(other.offered)
+            override fun hashCode() = 31 * pinned.contentHashCode() + offered.contentHashCode()
+        }
+    }
+
+    /**
+     * Compare what the relay just published against what is pinned.
+     *
+     * ⚠️ An empty or unparseable key counts as "not published" rather than as a
+     * change. A relay that answers with rubbish is broken, and treating broken
+     * as hostile would raise an alarm nobody can act on.
+     */
+    fun check(offeredBase64: String?, pinned: ByteArray?): Verdict {
+        val offered = offeredBase64
+            ?.takeIf { it.isNotEmpty() }
+            ?.let { runCatching { Wire.unb64(it) }.getOrNull() }
+            ?.takeIf { it.isNotEmpty() }
+            ?: return Verdict.NotPublished
+        if (pinned == null || pinned.isEmpty()) return Verdict.FirstSight(offered)
+        return if (offered.contentEquals(pinned)) Verdict.Unchanged
+        else Verdict.Changed(pinned = pinned, offered = offered)
+    }
+
+    /**
+     * What to tell somebody when it changed.
+     *
+     * 🔴 Names the two possibilities and does not pretend to know which. An
+     * alarm that asserts an attack when a rotation is equally likely trains
+     * people to dismiss alarms.
+     */
+    fun warning(host: String): String =
+        "The server at $host is presenting a different identity than the one " +
+            "this app recorded the first time it connected. That happens when the " +
+            "server's keys are rotated on purpose — and it also happens when " +
+            "something else is answering in its place. Nothing here can tell those " +
+            "apart. If you were not expecting a change, stop and check with " +
+            "whoever runs it before signing in again."
+}
+
+/**
+ * The relay is not the one this app pinned.
+ *
+ * 🔴 **Deliberately NOT a [KeyLogForkException].** A fork means the relay is
+ * showing two different histories and nothing it says can be trusted. This means
+ * the server's identity key is not the one recorded — a rotation somebody
+ * performed, or somebody else's server. The remedies are different, so the
+ * exceptions are different.
+ */
+class RelayIdentityChangedException(val host: String) :
+    RelayException(RelayIdentityPin.warning(host))

client/src/main/kotlin/io/flamenet/client/Trust.kt

@@ -612,21 +612,45 @@ class Trust(
     private fun verifyHeadSignature(head: Head, signatureB64: String) {
         if (signatureB64.isEmpty()) return
         val signature = runCatching { Wire.unb64(signatureB64) }.getOrNull() ?: return
-        var issuer = vault.keyLogIssuerPin()
-        if (issuer == null) {
-            val published = runCatching {
-                Wire.unb64(relay.relayDescriptor().optString("issuerPublicKey"))
-            }.getOrNull()
-            if (published == null || published.isEmpty()) return
-            vault.saveKeyLogIssuerPin(published)
-            issuer = published
+
+        // ios#19 — decide what the relay's published identity MEANS before
+        // using it, rather than inferring it from a signature that failed.
+        //
+        // 🔴 This is the bug the pin previously hid. A rotated relay key made
+        // the signature stop verifying, and the only conclusion available at
+        // that point was "fork" — so a command somebody ran on purpose was
+        // reported as the relay showing two different histories. Different
+        // situations, different remedies, and an alarm nobody can act on is one
+        // people learn to dismiss, which costs real fork detection too.
+        val offered = runCatching {
+            relay.relayDescriptor().optString("issuerPublicKey")
+        }.getOrNull()
+        val issuer = when (val verdict = RelayIdentityPin.check(offered, vault.keyLogIssuerPin())) {
+            is RelayIdentityPin.Verdict.NotPublished ->
+                // Nothing published and nothing pinned: a weaker deployment,
+                // not an attack. Nothing to check the signature against.
+                vault.keyLogIssuerPin() ?: return
+            is RelayIdentityPin.Verdict.FirstSight -> {
+                vault.saveKeyLogIssuerPin(verdict.offered)
+                verdict.offered
+            }
+            is RelayIdentityPin.Verdict.Unchanged -> vault.keyLogIssuerPin() ?: return
+            is RelayIdentityPin.Verdict.Changed ->
+                // ⚠️ NOT a fork, and `forkDetected` is deliberately not set.
+                // The log may be perfectly consistent; it is the server's
+                // identity that is not the one recorded.
+                throw RelayIdentityChangedException(relayHostForPin())
         }
+
         if (!KeyLog.verifyHead(head.size, head.root, signature, issuer)) {
             forkDetected = true
             throw KeyLogForkException("log head signature does not verify under the pinned key")
         }
     }
 
+    /** The host to name in an identity-change warning. */
+    private fun relayHostForPin(): String = relay.host
+
     private fun consistent(from: Int, to: Int, oldRoot: String, newRoot: String): Boolean {
         val c = relay.logConsistency(from.toLong(), to.toLong())
         return KeyLog.verifyConsistency(

client/src/test/kotlin/io/flamenet/client/RelayIdentityPinTest.kt

@@ -0,0 +1,167 @@
+package io.flamenet.client
+
+import java.util.Base64
+import kotlin.test.Test
+import kotlin.test.assertEquals
+import kotlin.test.assertIs
+import kotlin.test.assertTrue
+
+/**
+ * Pinning the relay's identity (ios#19, Android half).
+ *
+ * 🔴 **The finding this closes is not a missing feature — it is a wrong
+ * conclusion.** Android already pinned `issuerPublicKey`, inside
+ * `Trust.verifyHeadSignature`. When the key later changed, the head signature
+ * stopped verifying and the only conclusion available at that point was
+ * `KeyLogForkException` — so a rotation somebody performed on purpose was
+ * reported to the user as the relay showing two different histories.
+ *
+ * Those need different responses. A fork means nothing the relay says can be
+ * trusted. A rotation means somebody ran a command, or somebody else is
+ * answering on that address, and a person has to decide which. Reporting the
+ * second as the first is alarming and unactionable, and an alarm people cannot
+ * act on is one they learn to dismiss — which costs the real fork detection its
+ * credibility too.
+ */
+class RelayIdentityPinTest {
+
+    private fun key(byte: Int) = ByteArray(32) { byte.toByte() }
+    private fun b64(bytes: ByteArray): String = Base64.getEncoder().encodeToString(bytes)
+
+    // ---- The four verdicts ----------------------------------------------
+
+    @Test
+    fun `a published key with nothing pinned is first sight`() {
+        val offered = key(1)
+        val verdict = RelayIdentityPin.check(b64(offered), pinned = null)
+        assertEquals(RelayIdentityPin.Verdict.FirstSight(offered), verdict)
+    }
+
+    @Test
+    fun `the same key is unchanged`() {
+        assertEquals(
+            RelayIdentityPin.Verdict.Unchanged,
+            RelayIdentityPin.check(b64(key(1)), pinned = key(1)),
+        )
+    }
+
+    /**
+     * 🔴 The case the whole file exists for, and the one that used to be
+     * reported as a forked log.
+     */
+    @Test
+    fun `a different key is a change, carrying both halves`() {
+        val verdict = RelayIdentityPin.check(b64(key(2)), pinned = key(1))
+        val changed = assertIs<RelayIdentityPin.Verdict.Changed>(verdict)
+        assertTrue(changed.pinned.contentEquals(key(1)))
+        assertTrue(changed.offered.contentEquals(key(2)))
+    }
+
+    /**
+     * ⚠️ **A relay that publishes no key is a weaker deployment, not an
+     * attack.** Somebody self-hosting an older build has nothing to pin, and
+     * refusing to talk to them would be a worse failure than the one this
+     * prevents.
+     */
+    @Test
+    fun `a deployment that publishes nothing is not an attack`() {
+        assertEquals(
+            RelayIdentityPin.Verdict.NotPublished,
+            RelayIdentityPin.check(null, pinned = key(1)),
+        )
+        assertEquals(
+            RelayIdentityPin.Verdict.NotPublished,
+            RelayIdentityPin.check("", pinned = key(1)),
+        )
+    }
+
+    /**
+     * ⚠️ **Broken is not hostile.** A relay answering with rubbish counts as
+     * "not published" rather than as a change — treating a malformed field as
+     * an attack would raise an alarm nobody can act on, on a deployment that is
+     * simply misconfigured.
+     */
+    @Test
+    fun `an unparseable key reads as not published, not as a change`() {
+        assertEquals(
+            RelayIdentityPin.Verdict.NotPublished,
+            RelayIdentityPin.check("!!! not base64 !!!", pinned = key(1)),
+        )
+        // Valid base64 of nothing is still nothing.
+        assertEquals(
+            RelayIdentityPin.Verdict.NotPublished,
+            RelayIdentityPin.check(b64(ByteArray(0)), pinned = key(1)),
+        )
+    }
+
+    /** An empty pin is no pin: first sight, not a change against zero bytes. */
+    @Test
+    fun `an empty pin is treated as no pin`() {
+        val offered = key(3)
+        assertEquals(
+            RelayIdentityPin.Verdict.FirstSight(offered),
+            RelayIdentityPin.check(b64(offered), pinned = ByteArray(0)),
+        )
+    }
+
+    // ---- What somebody is told ------------------------------------------
+
+    /**
+     * 🔴 The warning names BOTH possibilities and asserts neither. An alarm that
+     * claims an attack when a rotation is equally likely trains people to
+     * dismiss alarms — and the next one may be real.
+     */
+    @Test
+    fun `the warning does not pretend to know which it is`() {
+        val text = RelayIdentityPin.warning("relay.example.com")
+        assertTrue(text.contains("relay.example.com"), text)
+        assertTrue(text.contains("rotated on purpose"), text)
+        assertTrue(text.contains("something else is answering"), text)
+        assertTrue(text.contains("Nothing here can tell those apart"), text)
+    }
+
+    /**
+     * 🔴 **The exception is not a fork exception, and that separation is the
+     * point of this change.** A caller catching `KeyLogForkException` to decide
+     * "the log is untrustworthy" must not be handed a rotation.
+     */
+    @Test
+    fun `an identity change is not reported as a key-log fork`() {
+        // ⚠️ Held as the sealed base, deliberately. Typed as the concrete class
+        // the compiler folds `!is KeyLogForkException` to a constant and warns —
+        // which is the compiler proving the point statically, and also means the
+        // assertion would test nothing. Through the base type it is the runtime
+        // check a caller's `when` actually performs.
+        val thrown: RelayException = RelayIdentityChangedException("relay.example.com")
+        assertTrue(
+            thrown !is KeyLogForkException,
+            "a rotated relay key is being reported as a forked log",
+        )
+        val identity = assertIs<RelayIdentityChangedException>(thrown)
+        assertEquals("relay.example.com", identity.host)
+        assertTrue(identity.message!!.contains("rotated on purpose"))
+    }
+
+    // ---- Shape -----------------------------------------------------------
+
+    /**
+     * ⚠️ `ByteArray` has reference equality by default, so the verdicts carry
+     * hand-written `equals`. Without them every comparison in this suite would
+     * be false and the tests would have been written around it.
+     */
+    @Test
+    fun `verdicts compare by content, not by reference`() {
+        assertEquals(
+            RelayIdentityPin.Verdict.FirstSight(key(1)),
+            RelayIdentityPin.Verdict.FirstSight(key(1)),
+        )
+        assertEquals(
+            RelayIdentityPin.Verdict.Changed(key(1), key(2)),
+            RelayIdentityPin.Verdict.Changed(key(1), key(2)),
+        )
+        assertTrue(
+            RelayIdentityPin.Verdict.Changed(key(1), key(2)) !=
+                RelayIdentityPin.Verdict.Changed(key(1), key(3)),
+        )
+    }
+}

engine/src/main/kotlin/io/flamenet/e2e/AccountKdf.kt

@@ -133,12 +133,17 @@ object AccountKdf {
         context: String,
         length: Int = 32,
         params: Params = Params.PBKDF2_V1,
-    ): ByteArray = Hkdf.derive(
-        ikm = master(passphrase, salt, params),
-        salt = ByteArray(0),
-        info = context.toByteArray(Charsets.UTF_8),
-        length = length,
-    )
+    ): ByteArray = SecureBytes.withSecret(master(passphrase, salt, params)) { ikm ->
+        // ios#39 — the stretched master is an intermediate nobody needs after
+        // the expansion, and it is the most valuable thing in this file: every
+        // other secret is derived from it. Wiped however this leaves.
+        Hkdf.derive(
+            ikm = ikm,
+            salt = ByteArray(0),
+            info = context.toByteArray(Charsets.UTF_8),
+            length = length,
+        )
+    }
 
     /** What gets sent to the account service. */
     fun authSecret(passphrase: String, salt: ByteArray, params: Params = Params.PBKDF2_V1): ByteArray =
@@ -179,14 +184,19 @@ object AccountKdf {
         salt: ByteArray,
         params: Params = Params.PBKDF2_V1,
     ): Pair<ByteArray, ByteArray> {
-        val ikm = master(passphrase, salt, params)
-        fun expand(context: String) = Hkdf.derive(
-            ikm = ikm,
-            salt = ByteArray(0),
-            info = context.toByteArray(Charsets.UTF_8),
-            length = 32,
-        )
-        return expand(Context.ACCOUNT_AUTH) to expand(Context.ACCOUNT_IDENTITY)
+        return SecureBytes.withSecret(master(passphrase, salt, params)) { ikm ->
+            fun expand(context: String) = Hkdf.derive(
+                ikm = ikm,
+                salt = ByteArray(0),
+                info = context.toByteArray(Charsets.UTF_8),
+                length = 32,
+            )
+            // ⚠️ BOTH expansions happen inside the block. Returning the pair and
+            // wiping afterwards would be the same thing; expanding after the
+            // wipe would derive two keys from thirty-two zero bytes, which is
+            // the failure mode this shape makes impossible to write.
+            expand(Context.ACCOUNT_AUTH) to expand(Context.ACCOUNT_IDENTITY)
+        }
     }
 
     /**

engine/src/main/kotlin/io/flamenet/e2e/SecureBytes.kt

@@ -0,0 +1,78 @@
+package io.flamenet.e2e
+
+import java.util.Arrays
+
+/**
+ * Zeroing key material once it has been used (ios#39).
+ *
+ * 🔴 **This narrows a window. It does not remove the material, and claiming
+ * otherwise would spend the whole value of the feature on a lie.** What it buys
+ * is that a derived key does not sit in the process for the lifetime of the app
+ * after the one moment it was needed — which matters for a heap dump, a crash
+ * report, or a debugger attached later. It buys nothing against somebody
+ * reading memory *while* the secret is in use.
+ *
+ * 🔴 **The JVM guarantee is weaker than the iOS one, and the difference is not
+ * a detail.** iOS has `memset_s`, which is specified not to be optimised away,
+ * and iOS does not page application memory to disk. Neither holds here:
+ *
+ * - **A compacting garbage collector moves objects.** A `ByteArray` can be
+ *   relocated between allocation and wiping, and the bytes at the old address
+ *   are not cleared by the move. Zeroing reaches the array you are holding; it
+ *   cannot reach a copy the runtime made on its own. There is no way to pin a
+ *   Java array, so this is a ceiling on what any implementation here can offer.
+ * - **Android has compressed swap.** zram has been on by default for years, so
+ *   a page holding key material can reach a backing store — unlike iOS, where
+ *   the platform does not page application memory out at all.
+ * - **There is no `memset_s`.** `Arrays.fill` on an array that escapes the
+ *   method is not eliminated in practice, because the JIT cannot prove the
+ *   write is dead. "Not in practice" is a weaker statement than the C standard
+ *   makes, and it is the honest one.
+ *
+ * ⚠️ **A Kotlin `String` cannot be zeroed**, same as Swift and for the same
+ * reasons: immutable, freely copied, and possibly interned. A passphrase that
+ * arrived as a `String` is beyond reach. The right response is to turn it into
+ * bytes early, wipe those, and not pretend about the original — which is why
+ * `AccountKdf.master` already calls `PBEKeySpec.clearPassword()` rather than
+ * trying to do anything about the `String` it was handed.
+ *
+ * So the claim this type supports, precisely: **a derived key is not left
+ * sitting in the heap after use.** Not "the key is gone".
+ */
+object SecureBytes {
+
+    /**
+     * Zero this array in place.
+     *
+     * ⚠️ Reaches THIS array only. Anything that copied it beforehand keeps an
+     * untouched copy, and the runtime may have made copies nobody asked for.
+     */
+    fun wipe(bytes: ByteArray?) {
+        if (bytes == null || bytes.isEmpty()) return
+        Arrays.fill(bytes, 0)
+    }
+
+    /**
+     * Run [body] with key material and wipe it however [body] leaves.
+     *
+     * ⚠️ Wipes on the way out of a throw as well. A derivation that failed
+     * half-way is exactly the case where the material is most likely to be
+     * forgotten — which is why this is `try/finally` rather than a wipe at the
+     * end of a happy path.
+     */
+    inline fun <T> withSecret(secret: ByteArray, body: (ByteArray) -> T): T =
+        try {
+            body(secret)
+        } finally {
+            wipe(secret)
+        }
+
+    /**
+     * A copy of [source], for callers that genuinely need one.
+     *
+     * 🔴 The copy is ordinary memory and this object stops being responsible
+     * for it. Every use of this is a decision to leave a second copy behind,
+     * and it is spelled out rather than implicit for that reason.
+     */
+    fun copyOf(source: ByteArray): ByteArray = source.copyOf()
+}

engine/src/test/kotlin/io/flamenet/e2e/SecureBytesTest.kt

@@ -0,0 +1,140 @@
+package io.flamenet.e2e
+
+import kotlin.test.Test
+import kotlin.test.assertContentEquals
+import kotlin.test.assertEquals
+import kotlin.test.assertFailsWith
+import kotlin.test.assertFalse
+import kotlin.test.assertTrue
+
+/**
+ * Zeroing key material once it has been used (ios#39, Android half).
+ *
+ * 🔴 **The claim being tested is narrow, and the tests are shaped to keep it
+ * narrow.** What this buys is that a derived key is not left sitting in the
+ * heap after the one moment it was needed. It is not "the key is gone" — a
+ * compacting collector can have moved the array, Android has compressed swap,
+ * and there is no `memset_s` on the JVM. `SecureBytes` says all of that in
+ * prose; these assert only what is actually true.
+ *
+ * 🔴 **The most valuable test here is that the derived keys did not change.**
+ * Wiping an intermediate one line too early derives every account secret from
+ * thirty-two zero bytes — which produces a perfectly consistent, perfectly
+ * wrong identity, silently, on every device. That is a far worse failure than
+ * the leak this feature exists to narrow, so it is checked first.
+ */
+class SecureBytesTest {
+
+    private val passphrase = "a-long-enough-passphrase"
+    private val salt = ByteArray(16) { it.toByte() }
+
+    // ---- The failure that would matter most -----------------------------
+
+    /**
+     * 🔴 The pinned vectors already cover `authSecret` and `accountKeySeed`.
+     * This covers the shape wiping could break without moving a vector: that
+     * the two secrets from one stretch still match the two derived separately.
+     * If the wipe landed between the expansions, the second would be derived
+     * from zeroes and this diverges.
+     */
+    @Test
+    fun `wiping the intermediate did not change what is derived`() {
+        val (auth, account) = AccountKdf.signInSecrets(passphrase, salt)
+        assertContentEquals(AccountKdf.authSecret(passphrase, salt), auth)
+        assertContentEquals(AccountKdf.accountKeySeed(passphrase, salt), account)
+    }
+
+    /**
+     * ⚠️ And neither is zeroes, which is what "derived from a wiped buffer"
+     * would look like. A test that only compared the two paths would pass if
+     * BOTH were wrong in the same way.
+     */
+    @Test
+    fun `no derived secret is a run of zeroes`() {
+        val (auth, account) = AccountKdf.signInSecrets(passphrase, salt)
+        assertFalse(auth.all { it == 0.toByte() }, "the auth secret is all zeroes")
+        assertFalse(account.all { it == 0.toByte() }, "the account seed is all zeroes")
+        assertFalse(
+            AccountKdf.identityKey(passphrase, salt).all { it == 0.toByte() },
+            "the identity key is all zeroes",
+        )
+    }
+
+    /** Derivation stays deterministic across calls, wiping or not. */
+    @Test
+    fun `derivation is still deterministic`() {
+        assertContentEquals(
+            AccountKdf.identityKey(passphrase, salt),
+            AccountKdf.identityKey(passphrase, salt),
+        )
+    }
+
+    /** Different contexts still separate, so the wipe did not collapse them. */
+    @Test
+    fun `the contexts still produce different secrets`() {
+        val auth = AccountKdf.authSecret(passphrase, salt)
+        val identity = AccountKdf.identityKey(passphrase, salt)
+        val account = AccountKdf.accountKeySeed(passphrase, salt)
+        assertFalse(auth.contentEquals(identity))
+        assertFalse(auth.contentEquals(account))
+        assertFalse(identity.contentEquals(account))
+    }
+
+    // ---- The wiping itself ----------------------------------------------
+
+    @Test
+    fun `wipe zeroes the array it is given`() {
+        val secret = ByteArray(32) { 0x41 }
+        SecureBytes.wipe(secret)
+        assertTrue(secret.all { it == 0.toByte() })
+    }
+
+    @Test
+    fun `wiping nothing is not an error`() {
+        SecureBytes.wipe(null)
+        SecureBytes.wipe(ByteArray(0))
+    }
+
+    @Test
+    fun `withSecret wipes on the way out`() {
+        val secret = ByteArray(32) { 0x41 }
+        val seen = SecureBytes.withSecret(secret) { it.copyOf() }
+        assertTrue(seen.all { it == 0x41.toByte() }, "the body saw a wiped buffer")
+        assertTrue(secret.all { it == 0.toByte() }, "the buffer was not wiped")
+    }
+
+    /**
+     * 🔴 **The case the `finally` exists for.** A derivation that failed
+     * half-way is exactly where the material gets forgotten, so the wipe has to
+     * survive a throw — not sit at the end of a happy path.
+     */
+    @Test
+    fun `withSecret wipes when the body throws`() {
+        val secret = ByteArray(32) { 0x41 }
+        assertFailsWith<IllegalStateException> {
+            SecureBytes.withSecret(secret) { error("derivation failed half-way") }
+        }
+        assertTrue(secret.all { it == 0.toByte() }, "a throw left the material behind")
+    }
+
+    /**
+     * ⚠️ **The honest limit, asserted rather than left to the prose.** A copy
+     * taken before the wipe is untouched by it. This is not a bug to fix — a
+     * `ByteArray` cannot reach copies of itself — and stating it as a passing
+     * test stops anybody reading `wipe` as stronger than it is.
+     */
+    @Test
+    fun `a copy taken beforehand survives the wipe`() {
+        val secret = ByteArray(32) { 0x41 }
+        val copy = SecureBytes.copyOf(secret)
+        SecureBytes.wipe(secret)
+        assertTrue(secret.all { it == 0.toByte() })
+        assertTrue(copy.all { it == 0x41.toByte() },
+            "the copy was wiped too, which would mean these share a buffer")
+    }
+
+    @Test
+    fun `withSecret returns what the body returned`() {
+        assertEquals(7, SecureBytes.withSecret(ByteArray(4)) { 7 })
+    }
+}