@cello-protocol/daemon 0.0.169 → 0.0.171

package/dist/session-node-manager.js

@@ -52,6 +52,74 @@ import { GATEWAY_UNAVAILABLE, GOVERNANCE_TIMEOUT, } from "@cello-protocol/gatewa
  */
 const AUTOACK_BROKER_GRACE_MS = 30_000;
 const MAX_REFUSED_PARKED_ENTRIES = 512;
+/**
+ * DOD-M12B-ACK-1 — inbound `/cello/content/1.0.0` streams allowed per connection.
+ *
+ * libp2p's registrar default is 32, and it enforces the cap AFTER protocol negotiation has already
+ * answered, so exceeding it resets a stream the sender believes it just opened and the sender's
+ * next write fails with a message that names nothing. Content delivery is bursty by design (a
+ * document sweep opened 99 events in one second on a live daemon), and a slot stays occupied for
+ * the whole of ingest — which awaits SQLCipher and the security gateway. 32 is simply too close to
+ * normal traffic to be a safety limit.
+ *
+ * This is headroom, NOT the fix. The fix is that #handleContentStream now closes what it opens; a
+ * raised cap without that would only move the cliff. Kept finite on purpose: an unbounded cap would
+ * let a peer pin memory by opening streams it never uses, and the ceiling is what makes a future
+ * leak of this shape show up as a bounded failure instead of a heap.
+ */
+const CONTENT_MAX_INBOUND_STREAMS = 512;
+/**
+ * DOD-M12B-ACK-1 — how long an inbound content stream may stay open after we have closed our end.
+ *
+ * Closing our write end retires the stream only once the PEER has closed its end too, so a peer
+ * that opens streams and never closes them still fills our inbound slots — a guard that runs only
+ * on the party it constrains is not a guard. After this window we reset it ourselves, which the
+ * muxer honours unilaterally.
+ *
+ * It cannot be zero: an immediate reset would land while a well-behaved sender is still inside its
+ * own `await stream.close()`, rejecting that close and turning every ordinary send into a park.
+ * The frame is already ingested long before this fires, so nothing waits on it.
+ */
+const CONTENT_STREAM_LINGER_MS = 30_000;
+/**
+ * DOD-M12B-SHUTDOWN-1 — how long one teardown step may block the daemon's exit.
+ *
+ * Chosen against the surface that complains: `cello logout` gives up and reports the daemon still
+ * running after 5 s, so a step that can burn longer than that guarantees the message the operator
+ * saw. Two steps at 2 s each stay inside it.
+ */
+const SHUTDOWN_STEP_DEADLINE_MS = 2_000;
+/**
+ * DOD-M12B-REDIAL-1 — the shortest gap between two re-dials of one session.
+ *
+ * Long enough that a burst of sends against a peer that is genuinely gone costs one dial rather
+ * than one per message; short enough that a peer coming back is picked up on the next thing the
+ * operator says. It is cleared on a successful dial, so it never delays a live counterparty.
+ */
+const REDIAL_COOLDOWN_MS = 15_000;
+/**
+ * DOD-CAP-SELF-HEAL-1 — what counts against a per-sender acceptance bound.
+ *
+ * `active` always. `interrupted` ONLY when the counterparty caused it.
+ *
+ * D18 is why `interrupted` has to count at all: a peer can flip a session to `interrupted` for free
+ * by dropping its stream, then open a fresh one, indefinitely. Those are theirs and still count.
+ *
+ * What broke was charging them for OURS. A daemon restart flips every live session to
+ * `interrupted`, nothing resolves them, and the reaper correctly refuses to take any with received
+ * content — so the bound became all-time instead of concurrent. Measured 2026-08-17: two of one
+ * operator's own agents could not open a session, because one held five finished conversations with
+ * the other against a stranger cap of three.
+ *
+ * NULL counts as the counterparty's. The column is new, so every pre-existing row is unlabelled,
+ * and the safe default for an anti-abuse bound is to count rather than to excuse.
+ */
+const CAP_COUNTS = (alias = "") => {
+    const p = alias ? `${alias}.` : "";
+    return `(${p}status = 'active'
+           OR (${p}status = 'interrupted' AND COALESCE(${p}interrupted_by, 'counterparty') != 'local'))`;
+};
+const CAP_COUNT_SQL = (where) => `SELECT COUNT(*) AS n FROM sessions WHERE ${where} AND ${CAP_COUNTS()}`;
 // Persistence bounds are TIER-GRADUATED via DEFAULT_TIER_BOUNDS (contacts-tier-migration). The two
 // consts below DERIVE from the grid's UNKNOWN row rather than restating it — the grid is the single
 // source (DOD-TIER-2 AC4), so these can never drift from it.
@@ -237,9 +305,36 @@ export class SessionNodeManager {
     #autoNatProbers;
     // M7-SESSION-003: per-session direct-path counterparty liveness, observed on the
     // session node's onPeerConnect ('alive') / onPeerDisconnect ('gone'). This is
-    // the liveness authority for direct sessions — the unilateral-seal gate reads
-    // it (relay sessions query the relay instead). NEVER the directory (SI-002).
+    // the liveness authority for direct sessions (relay sessions query the relay
+    // instead). NEVER the directory (SI-002). Read by exactly three consumers: the
+    // half-open reaper, both status surfaces, and cello_receive. No seal path reads
+    // it — the coupling to sealing runs through the receive guidance, which turns
+    // 'gone' into "call cello_close_session".
+    // DOD-M12B-ACK-1 adds 'impaired': connection up, our writes on it failing. It sits BELOW
+    // 'alive' and never above 'gone' — see #markSessionImpaired.
     #sessionLiveness = new Map();
+    // DOD-M12B-ACK-1: WHY a session is impaired and what became of the content. Separate from the
+    // state above because the state is what surfaces print and this is what they must explain.
+    #impairmentCause = new Map();
+    // DOD-M12B-STRAND-1: sessions whose durable holds have been read back. One read per session per
+    // process; the Map is the working copy from then on.
+    #heldRestored = new Set();
+    // DOD-M12B-SEAL-STUCK-1: sessions whose post-restore release has been attempted. Separate from
+    // #heldRestored so a READ-ONLY probe can hydrate without performing (or consuming) the release.
+    #heldReleased = new Set();
+    // DOD-M12B-SEAL-STUCK-1: sessions whose ordering state THIS PROCESS has observed — a relay
+    // witness recorded for it. `#witnessedSeq` is memory-only, so for a session that predates this
+    // daemon "no gap recorded" means "not recorded", not "no gap", and that difference decides
+    // whether we may tell an operator the session is safe to close.
+    #orderingObserved = new Set();
+    // DOD-M12B-INDEX-1: sessions whose tree and whose relay counter have provably parted. A diverged
+    // session can never produce a root the counterparty agrees with, so it must never be reported as
+    // safe to close — the close would be signed, refused as `leaf_count_mismatch`, and the receipt
+    // lost for good.
+    #diverged = new Set();
+    // DOD-M12B-ACK-1: pending linger-resets for inbound content streams the peer has not closed.
+    // Held so shutdown can drop them rather than leave timers pointing at a torn-down node.
+    #lingeringStreams = new Set();
     // M7-UPGRADE-002: sessions whose content integrity could NOT be verified (a content_hash
     // mismatch = tamper was observed). The auto-acknowledge gate (SI-002) refuses to auto-co-sign
     // for a desynced session — B must never blind-sign a tail it cannot verify. Keyed by sessionId hex.
@@ -341,6 +436,20 @@ export class SessionNodeManager {
     // with its agent away), and the park deposit that follows has to be refused. One without the
     // other reproduces nothing.
     #sendFaultRemaining = 0;
+    // DOD-M12B-ACK-1 — the same seam for the delivery-ACK write. See injectAckFault.
+    #ackFaultRemaining = 0;
+    // DOD-M12B-REDIAL-1 — makes the next N `newStream` calls report the connection as gone, BEFORE
+    // the node is touched. The sibling of injectSendFault for the one condition that used to end a
+    // conversation permanently; a real connection drop is not reproducible in-process.
+    #connectionLossRemaining = 0;
+    // DOD-M12B-REDIAL-1: the counterparty addresses this session dialled, kept so it can dial them
+    // again. They arrived in the FROST-signed assignment and were used once and dropped, which is
+    // why nothing could ever re-dial.
+    #counterpartyAddrs = new Map();
+    // DOD-M12B-REDIAL-1: when this session may next attempt a re-dial. Continuous re-dialling is what
+    // produced the 2026-08-17 notification storm, so a burst of sends against a peer that is gone
+    // costs one attempt, not one per message.
+    #redialNotBefore = new Map();
     /** Arm the park-deposit fault. Returns the count now armed. */
     injectParkFault(count, cause) {
         this.#parkFaultRemaining = Math.max(0, count);
@@ -353,6 +462,20 @@ export class SessionNodeManager {
         this.#sendFaultRemaining = Math.max(0, count);
         return this.#sendFaultRemaining;
     }
+    /** DOD-M12B-ACK-1: arm the delivery-ACK write fault — the sibling of injectSendFault for the
+     *  path that fails on a LISTENING agent. Without it the ACK failure branch (which impairs the
+     *  session and, until this milestone, could never clear it again) is unreachable from a test:
+     *  a listener sends no content, so the direct-send fault never fires for it. */
+    injectAckFault(count) {
+        this.#ackFaultRemaining = Math.max(0, count);
+        return this.#ackFaultRemaining;
+    }
+    /** DOD-M12B-REDIAL-1: arm the connection-loss fault — the next N direct sends find no open
+     *  connection, exactly as they do after any blip. See #connectionLossRemaining. */
+    injectConnectionLoss(count) {
+        this.#connectionLossRemaining = Math.max(0, count);
+        return this.#connectionLossRemaining;
+    }
     getSendFaultRemaining() {
         return this.#sendFaultRemaining;
     }
@@ -639,6 +762,17 @@ export class SessionNodeManager {
             // watermark: this records "operator acknowledged via dismiss", not "operator received via
             // cello_receive". NULL = not yet dismissed.
             "ALTER TABLE sessions ADD COLUMN read_at INTEGER",
+            // DOD-M12B-ABANDON-NOTIFY-1: epoch-ms when the counterparty told us they force-abandoned.
+            // Deliberately NOT a status — the session stays sealable, so the operator can still take a
+            // unilateral receipt. It stops this side calling them, nothing more.
+            "ALTER TABLE sessions ADD COLUMN counterparty_abandoned_at INTEGER",
+            // DOD-CAP-SELF-HEAL-1: WHO caused this session to be interrupted — 'counterparty' when their
+            // stream dropped, 'local' when OUR daemon stopped or started. Only theirs counts against the
+            // acceptance bound. Without this the bound is all-time rather than concurrent: every restart
+            // flips every live session to `interrupted`, nothing ever resolves them, and a pair of agents
+            // that has talked three times can never talk again. NULL means "not recorded" and is treated
+            // as the counterparty's, because the safe default for an anti-abuse bound is to count it.
+            "ALTER TABLE sessions ADD COLUMN interrupted_by TEXT",
         ]) {
             try {
                 this.#db.exec(ddl);
@@ -733,6 +867,69 @@ export class SessionNodeManager {
         PRIMARY KEY (agent_id, session_id, leaf_index)
       )
     `);
+        // DOD-M12B-STRAND-1 — content we RECEIVED and VERIFIED but cannot append yet.
+        //
+        // Held content used to live only in `#heldContent`, a Map that died with the session node. The
+        // teardown path said so itself: "the content is unrecoverable by the time we are here."
+        // Measured on one daemon in one morning: 367 held, 8 released, **24 destroyed**. Each
+        // destruction is permanent and one-sided — the sender was never acknowledged, so it believes
+        // the message is merely pending, while the only copy the receiver will ever see is gone and
+        // every later message in that session is stuck behind a gap nothing can fill.
+        //
+        // `canonical_seq` is the RELAY's position, not a local counter, and it is part of the key: that
+        // is what lets a frame come back after a restart and land at its OWN index rather than the next
+        // free slot. Appending it anywhere else would change the root the seal signs over.
+        //
+        // Keyed on agent_id, never agent_name — agent_name is a mutable display label (see the repo
+        // guide). `content_blob` is the SCREENED copy that gets delivered; `original_blob` is the peer's
+        // raw bytes, which the release path needs because classification reads byte 0 and the screened
+        // copy is no longer a CBOR map header for a document frame.
+        this.#db.exec(`
+      CREATE TABLE IF NOT EXISTS held_content (
+        agent_id TEXT NOT NULL,
+        session_id TEXT NOT NULL,
+        canonical_seq INTEGER NOT NULL,
+        content_blob BLOB NOT NULL,
+        original_blob BLOB,
+        content_hash_hex TEXT NOT NULL,
+        screened_out INTEGER NOT NULL DEFAULT 0,
+        correlation_id TEXT,
+        held_at INTEGER NOT NULL,
+        -- DOD-M12B-INDEX-1: 'received' (default) or 'sent'. A held frame of OUR OWN must be
+        -- released down the sent path — appended and transcribed as sent — never down the received
+        -- path, which would put our words in the counterparty's mouth in the sealed record and hand
+        -- them back to our own agent through cello_receive as though they had just arrived.
+        origin TEXT NOT NULL DEFAULT 'received',
+        -- DOD-M12B-INDEX-1: 'msg' or 'doc'. A held document leaf must come back as a document leaf.
+        leaf_kind TEXT NOT NULL DEFAULT 'msg',
+        PRIMARY KEY (agent_id, session_id, canonical_seq)
+      )
+    `);
+        // DOD-M12B-INDEX-1: `CREATE TABLE IF NOT EXISTS` is a NO-OP against a table that already
+        // exists, so a database created between DOD-M12B-STRAND-1 and this change has `held_content`
+        // WITHOUT `origin`. On those every insert throws and every restore throws — holds go back to
+        // memory-only, silently at the surface, and that now includes our own sent messages, which
+        // nobody else holds a copy of. Loud in the log is not the same as visible.
+        try {
+            this.#db.exec("ALTER TABLE held_content ADD COLUMN origin TEXT NOT NULL DEFAULT 'received'");
+        }
+        catch (err) {
+            const msg = err instanceof Error ? err.message : String(err);
+            if (!/duplicate column name/i.test(msg))
+                throw err;
+        }
+        // DOD-M12B-INDEX-1: and the LEAF KIND. `#releaseHeld` used to append every held frame as "msg",
+        // so a document leaf that had to wait for its position came back as a conversation message —
+        // the distinction survived the immediate append and was destroyed by the hold, unrecoverably
+        // after a restart.
+        try {
+            this.#db.exec("ALTER TABLE held_content ADD COLUMN leaf_kind TEXT NOT NULL DEFAULT 'msg'");
+        }
+        catch (err) {
+            const msg = err instanceof Error ? err.message : String(err);
+            if (!/duplicate column name/i.test(msg))
+                throw err;
+        }
         // DOD-LOG-1 (PERSIST-LOG-001) / PERSIST-002 (AC-010): the durable, ENCRYPTED-at-rest readable
         // transcript. Each row is keyed by the canonical leaf `sequence`, so it JOINS to
         // session_tree_leaves(leaf_index) — a stored message is provably behind a committed hash-chain
@@ -890,7 +1087,10 @@ export class SessionNodeManager {
             for (const row of activeRows) {
                 try {
                     this.#db
-                        .prepare("UPDATE sessions SET status = 'interrupted', updated_at = ?, interrupted_at = COALESCE(interrupted_at, ?) WHERE agent_id = ? AND session_id = ?")
+                        .prepare(
+                    // DOD-CAP-SELF-HEAL-1: OURS. This is the boot sweep finding sessions a previous
+                    // process left `active`; the counterparty did nothing, so they are not charged for it.
+                    "UPDATE sessions SET status = 'interrupted', updated_at = ?, interrupted_at = COALESCE(interrupted_at, ?), interrupted_by = 'local' WHERE agent_id = ? AND session_id = ?")
                         .run(now, interruptedAt, row.agent_id, row.session_id);
                     if (row.agent_name === null) {
                         this.#logger.error("session.agent.orphaned", {
@@ -1548,6 +1748,10 @@ export class SessionNodeManager {
      *  otherwise let multiple held chunks each individually pass the size gate while cumulatively
      *  exceeding it once #releaseHeld drains them. */
     #getHeldBytesTotal(agentName, sessionId) {
+        // DOD-M12B-STRAND-1: hydrate first. Reading the Map before the durable holds are back
+        // under-counts, and this gate exists to stop several held chunks each passing the size cap
+        // individually while cumulatively exceeding it — an under-count is the bypass.
+        this.#ensureHeldRestored(agentName, sessionId);
         const held = this.#heldContent.get(this.#k(agentName, sessionId));
         if (!held)
             return 0;
@@ -1567,7 +1771,7 @@ export class SessionNodeManager {
         if (!this.#db)
             return 0;
         const row = this.#db
-            .prepare("SELECT COUNT(*) AS n FROM sessions WHERE agent_id = ? AND counterparty_pubkey = ? AND status IN ('active', 'interrupted')")
+            .prepare(CAP_COUNT_SQL("agent_id = ? AND counterparty_pubkey = ?"))
             .get(this.#requireAgentId(agentName), counterpartyPubkey);
         return row.n;
     }
@@ -1583,7 +1787,8 @@ export class SessionNodeManager {
             return 0;
         const row = this.#db
             .prepare(`SELECT COUNT(*) AS n FROM sessions s
-         WHERE s.agent_id = ? AND s.status IN ('active', 'interrupted')
+         WHERE s.agent_id = ?
+           AND ${CAP_COUNTS("s")}
            AND NOT EXISTS (
              SELECT 1 FROM contacts c
              WHERE c.agent_id = s.agent_id AND c.pubkey = s.counterparty_pubkey
@@ -1605,6 +1810,10 @@ export class SessionNodeManager {
         const perSenderCap = this.resolveTierBound(agentName, tier, "max_sessions");
         const perSender = this.countActiveSessionsForCounterparty(agentName, counterpartyPubkey);
         if (perSender >= perSenderCap) {
+            // BYTE-IDENTICAL to every other refusal, deliberately — DOD-TIER-3. A BLOCKED sender and an
+            // over-cap UNKNOWN must be indistinguishable, or the refusal tells someone they are blocked.
+            // The operator's alarm needs numbers, so it asks for them SEPARATELY via capDiagnostics;
+            // hanging them off this object would put a distinguishing oracle in the return value.
             return { ok: false, reason: "abuse_bound_sessions_per_sender" };
         }
         // The global stranger cap is only for the UNKNOWN pool. A KNOWN+ sender is past it by trust;
@@ -1978,10 +2187,21 @@ export class SessionNodeManager {
             };
         }
         // Log observability event (session.node.created)
+        //
+        // `counterpartySessionPeerId` IS LOGGED because it is recorded here ONCE and never refreshed,
+        // while a standing receiver is rebuilt with a fresh libp2p keypair on every signaling reconnect
+        // and every lost reservation. If the peer rebuilds between advertising its endpoint and this
+        // handoff, we record an identity that no longer exists — and since `newStream` never dials, it
+        // only ever looks for an ALREADY-OPEN connection filed under exactly this string, so every send
+        // in this direction parks forever while the reverse direction works fine.
+        //
+        // Both sides of a local session log this event, so recording the id we will dial makes that
+        // mismatch a direct comparison in the log instead of an unfalsifiable hypothesis.
         this.#logger.info("session.node.created", {
             sessionId,
             agentName,
             sessionPeerId: peerId,
+            counterpartySessionPeerId: counterpartyPeerId,
             correlationId,
         });
         // Add to active map (keyed by (agentName, sessionId) — DOD-LOOP-1)
@@ -2232,10 +2452,176 @@ export class SessionNodeManager {
     /**
      * M7-SESSION-003: read the direct-path counterparty liveness for a session.
      * 'unknown' when no session node observation has occurred yet.
+     *
+     * DOD-M12B-ACK-1: 'impaired' is DAEMON-LOCAL and deliberately not on the relay's
+     * SessionLiveness wire type — the relay answers a different question (does it hold the
+     * recipient's standing connection) and its three states are a deployed bilateral contract.
      */
     getSessionLiveness(agentName, sessionId) {
         return this.#sessionLiveness.get(this.#k(agentName, sessionId)) ?? "unknown";
     }
+    /**
+     * DOD-M12B-ACK-1 — live `/cello/content/1.0.0` stream counts on a session's direct path, or null
+     * when the session has no active node.
+     *
+     * Answerable at runtime on purpose, in the same spirit as getConnectionMonitorPolicy: the count
+     * is what decides whether the next send survives, and until this existed it could only be
+     * recovered by measuring a log after the fact. It is also what lets the regression assert that a
+     * slot was RELEASED, rather than that some particular number of messages happened to fit.
+     */
+    countSessionContentStreams(agentName, sessionId) {
+        const entry = this.#activeNodes.get(this.#k(agentName, sessionId));
+        if (!entry || typeof entry.node.countProtocolStreams !== "function")
+            return null;
+        return entry.node.countProtocolStreams(entry.counterpartySessionPeerId, CELLO_CONTENT_PROTOCOL_ID);
+    }
+    /**
+     * DOD-M12B-ACK-1 — the live content-stream counts for a peer, as log context.
+     *
+     * A diagnostic must not be able to break the failure path it describes, so a node that predates
+     * `countProtocolStreams` (test fakes do) yields no fields rather than throwing.
+     */
+    #streamCensus(node, peerId) {
+        if (typeof node.countProtocolStreams !== "function")
+            return {};
+        try {
+            const { inbound, outbound } = node.countProtocolStreams(peerId, CELLO_CONTENT_PROTOCOL_ID);
+            return { contentStreamsInbound: inbound, contentStreamsOutbound: outbound, contentStreamsInboundCap: CONTENT_MAX_INBOUND_STREAMS };
+        }
+        catch {
+            return {};
+        }
+    }
+    /**
+     * DOD-M12B-ACK-1 — the connection is up and delivery on it is not working.
+     *
+     * Liveness is otherwise driven ONLY by libp2p peer-connect/peer-disconnect, so it answers "is
+     * there a connection object?" while every surface that prints it is read as "can I talk to them?".
+     * Measured 2026-08-17: one session reported `alive` for 70 minutes after every write had started
+     * failing, another never stopped.
+     *
+     * ONLY 'gone' is protected, and 'gone' is NOT protected because a seal gate reads it — nothing in
+     * the code does. It is protected because the receive surface turns 'gone' into "call
+     * cello_close_session", and a failed write must never be able to produce that instruction.
+     *
+     * 'unknown' is DOWNGRADED just like 'alive', which is not obvious and is the point. A session
+     * whose recorded `counterpartySessionPeerId` has gone stale never sees a matching peer-connect,
+     * so it sits at 'unknown' while every send fails forever — the exact case documented at
+     * #wireSessionLiveness — and the receive surface renders 'unknown' as healthy-and-quiet, which is
+     * the 70-minute lie relocated one lane over. 'unknown' claims nothing; the surface built on it does.
+     */
+    #markSessionImpaired(agentName, sessionId, opts) {
+        const key = this.#k(agentName, sessionId);
+        const prior = this.#sessionLiveness.get(key);
+        if (prior === "gone") {
+            // Declining is a decision, so it is logged. A silent early return here is the shape that let
+            // the original defect hide for a day: nothing recorded that writes were failing on a session
+            // every surface was still calling healthy.
+            this.#logger.debug("session.liveness.impairment.declined", {
+                sessionId, liveness: prior, cause: opts.cause, error: opts.error, correlationId: opts.correlationId,
+            });
+            return;
+        }
+        // The CAUSE is refreshed even when the state does not move, because the receive surface builds
+        // its guidance from it and a stale cause would describe the wrong failure.
+        this.#impairmentCause.set(key, { cause: opts.cause, retained: "unknown" });
+        if (prior === "impaired")
+            return;
+        this.#sessionLiveness.set(key, "impaired");
+        this.#logger.warn("session.liveness.changed", {
+            sessionId,
+            counterpartyPubkey: this.getSessionRecord(agentName, sessionId)?.counterparty_pubkey,
+            transportPath: "direct",
+            liveness: "impaired",
+            observedBy: opts.cause,
+            priorLiveness: prior ?? "unknown",
+            // `reason` is a contract string and `error` is the message — the convention every other
+            // failure log in this file follows. Collapsing them makes grouping by `reason` useless.
+            reason: "write_failed",
+            error: opts.error,
+            correlationId: opts.correlationId,
+        });
+    }
+    /**
+     * DOD-M12B-ACK-1 — what became of the content whose send caused the impairment.
+     *
+     * The receive surface has no memory of the last send, so without this it can only guess — and the
+     * guess it would make ("it was parked, do not resend") is FALSE in the two cases that matter
+     * most: a refused park whose durable enqueue was dropped, and one that threw. In both the message
+     * is gone and `cello_send` has already told the caller to send it again, so a receive that says
+     * "do not resend" contradicts it later, while the agent is sitting there waiting.
+     */
+    #noteImpairmentRetention(agentName, sessionId, retained) {
+        const key = this.#k(agentName, sessionId);
+        const current = this.#impairmentCause.get(key);
+        if (!current)
+            return;
+        this.#impairmentCause.set(key, { cause: current.cause, retained });
+    }
+    /** DOD-M12B-ACK-1: why this session is impaired, for the surface that has to explain it. Null
+     *  when it is not impaired — a caller must not narrate a failure that is not current. */
+    getSessionImpairment(agentName, sessionId) {
+        const key = this.#k(agentName, sessionId);
+        if (this.#sessionLiveness.get(key) !== "impaired")
+            return null;
+        return this.#impairmentCause.get(key) ?? null;
+    }
+    /**
+     * DOD-M12B-ACK-1 — a delivery landed, so the impairment is over.
+     *
+     * Without this an `impaired` flag is a one-way door: one bad write would make a session report a
+     * broken conversation for the rest of its life, which is the same class of lie in the other
+     * direction. Called from BOTH send paths — an agent that mostly listens sends content rarely and
+     * ACKs constantly, so clearing only on content would leave exactly those sessions impaired
+     * forever. Only clears 'impaired': a successful write says nothing about a connection libp2p has
+     * already declared 'gone'.
+     */
+    #clearSessionImpairment(agentName, sessionId, observedBy, correlationId) {
+        const key = this.#k(agentName, sessionId);
+        if (this.#sessionLiveness.get(key) !== "impaired")
+            return;
+        this.#sessionLiveness.set(key, "alive");
+        this.#impairmentCause.delete(key);
+        this.#logger.info("session.liveness.changed", {
+            sessionId,
+            counterpartyPubkey: this.getSessionRecord(agentName, sessionId)?.counterparty_pubkey,
+            transportPath: "direct",
+            liveness: "alive",
+            observedBy,
+            reason: "write_succeeded",
+            correlationId,
+        });
+    }
+    /**
+     * DOD-M12B-ABANDON-NOTIFY-1 — drive the REAL inbound content handler with one framed message and
+     * a claimed peer identity.
+     *
+     * The handler is registered on a live libp2p node, so without this the only way to reach its
+     * branches is a full two-node transport fixture — which is why the session-abandoned branch and
+     * its peer pinning had no coverage at all. This feeds the same function the protocol handler
+     * calls, including the authentication check, rather than a copy of its logic.
+     */
+    async handleContentFrameForTest(agentName, sessionId, framedBytes, remotePeerId) {
+        const source = {
+            async *[Symbol.asyncIterator]() { yield framedBytes; },
+            close: async () => { },
+            abort: () => { },
+            status: "closed",
+        };
+        await this.#handleContentStream(agentName, sessionId, source, remotePeerId);
+    }
+    /** DOD-CAP-SELF-HEAL-1 test seam: the shutdown sweep's effect, without a shutdown. Mirrors the
+     *  real UPDATE at gracefulShutdown so a test cannot pass against a label production never sets. */
+    markSessionsInterruptedByLocalShutdownForTest() {
+        // SCOPED TO UNLABELLED ROWS. Unscoped, this would relabel rows already marked
+        // `counterparty` — one call excusing every interruption every attacker ever caused, on every
+        // agent. Production never relabels; nor does this.
+        this.#db?.prepare("UPDATE sessions SET interrupted_by = 'local' WHERE status = 'interrupted' AND interrupted_by IS NULL").run();
+    }
+    /** DOD-CAP-SELF-HEAL-1 test seam: the counterparty's stream closing, without a real peer. */
+    markInterruptedByCounterpartyForTest(agentName, sessionId) {
+        this.#db?.prepare("UPDATE sessions SET interrupted_by = 'counterparty' WHERE agent_id = ? AND session_id = ?").run(this.#requireAgentId(agentName), sessionId);
+    }
     /** Test seam (same spirit as getDb()): seed per-session direct-path liveness, which is otherwise
      *  only set by the live node's onPeerConnect/onPeerDisconnect (#wireSessionLiveness). Lets a
      *  DB-seeded test exercise the CC-5 reaper's "alive counterparty must survive" gate without standing
@@ -2308,11 +2694,16 @@ export class SessionNodeManager {
                     "Check the daemon's database (disk space, permissions).",
             };
         }
-        // Log observability event
+        // Log observability event. `counterpartySessionPeerId` for the same reason as the initiator
+        // side: this is the identity every later send will look for an open connection under, it is
+        // never refreshed, and the peer's standing receiver may already have been rebuilt under a new
+        // one. The RESPONDER is the side that can go stale — only the initiator dials, so this is the
+        // half that inherits an id it never verified.
         this.#logger.info("session.node.created", {
             sessionId,
             agentName,
             sessionPeerId: peerId,
+            counterpartySessionPeerId: initiatorPeerId,
             correlationId,
         });
         // Remove this agent's standing receiver from the slot and add to active map. The handed-off
@@ -2385,7 +2776,10 @@ export class SessionNodeManager {
         // surface as interrupted so AC-010 recovery handles them at next login.
         // The session.node.destroyed log preserves the original reason for observability.
         const dbStatus = reason === "sealed" ? "sealed" : "interrupted";
-        this.#updateSessionStatus(agentName, sessionId, dbStatus);
+        // DOD-CAP-SELF-HEAL-1: OURS. Every caller of this with a non-sealed reason is a local teardown
+        // — the operator's kill switch (`cello_set_agent_offline`), an internal error, a node replaced.
+        // The counterparty did nothing, so they must not be charged a cap slot for it.
+        this.#updateSessionStatus(agentName, sessionId, dbStatus, dbStatus === "interrupted" ? "local" : undefined);
         this.#activeNodes.delete(this.#k(agentName, sessionId));
         // Evict the in-memory per-session caches on teardown. The tree is durable in
         // SQLite (getSessionTree reloads it on demand), and the received-content buffer
@@ -2500,22 +2894,69 @@ export class SessionNodeManager {
         // destroyed with the session three seconds later. The sender then re-sent that envelope 90
         // times against a ceiling of 5, and every surface reported the delivery as merely pending.
         //
-        // This is a LOSS REPORT, not a fix — the content is unrecoverable by the time we are here. The
-        // fix is upstream, in not tearing a session down while an answer is still owed on it.
+        // DOD-M12B-STRAND-1 — THIS IS NO LONGER AN EPITAPH. Dropping the Map now drops a CACHE: the
+        // frames are rows in `held_content`, and #restoreHeldContent brings them back the next time
+        // this session gets a node. `session.content.held.discarded` is kept, at WARN rather than
+        // ERROR, and only for what it now means — a gap is still open on a session going away, which
+        // is worth an alarm even though nothing is lost.
+        //
+        // It fires ONLY when the durable rows are confirmed present. A hold whose persist failed is a
+        // genuine loss and must not be reported with the same event as a survivor, so it gets its own
+        // error naming the count that is actually gone. The check is a COUNT against the store rather
+        // than a belief about the write that ran earlier.
         const strandedHolds = this.#heldContent.get(key);
         if (strandedHolds && strandedHolds.size > 0) {
-            this.#logger.error("session.content.held.discarded", {
+            const canonicalSeqs = [...strandedHolds.keys()].sort((a, b) => a - b);
+            // NULL means "we could not find out", and it must never render as "destroyed". The bare
+            // catch this replaces coerced a failed COUNT to 0, which then claimed every held frame had
+            // been lost — asserting a cause it had not established. It is not hypothetical:
+            // #requireAgentId THROWS for a retired agent, on this exact path, so retiring an agent with
+            // an open hold fabricated a data-loss alarm pointing at the persistence layer while the real
+            // fault was name resolution.
+            let durable = null;
+            try {
+                const row = this.#db?.prepare("SELECT COUNT(*) AS n FROM held_content WHERE agent_id = ? AND session_id = ?").get(this.#requireAgentId(agentName), sessionId);
+                durable = row?.n ?? 0;
+            }
+            catch (err) {
+                this.#logger.warn("session.content.held.durable_count.failed", {
+                    agentName, sessionId,
+                    impact: "cannot say whether the held frames are durable — reported as unknown, NOT as lost",
+                    error: err instanceof Error ? err.message : String(err),
+                });
+            }
+            const lost = durable === null ? null : strandedHolds.size - durable;
+            this.#logger.warn("session.content.held.discarded", {
                 agentName,
                 sessionId,
                 count: strandedHolds.size,
-                canonicalSeqs: [...strandedHolds.keys()].sort((a, b) => a - b),
+                durable,
+                canonicalSeqs,
                 // NULL when the tree was not cached at teardown. Honest, and cheap — reloading the leaf
                 // table to fill in a diagnostic field is not worth a disk read on every teardown, let
                 // alone the cache resurrection it caused.
                 treeSize: treeSizeBeforeEviction,
             });
+            if (lost !== null && lost > 0) {
+                this.#logger.error("session.content.held.lost", {
+                    agentName,
+                    sessionId,
+                    lost,
+                    held: strandedHolds.size,
+                    impact: "verified content was NOT written to held_content and is destroyed by this teardown — the sender was never acknowledged and believes it is still pending",
+                });
+            }
         }
         this.#heldContent.delete(key);
+        // DOD-M12B-STRAND-1: the hydration guard goes with the cache it guards. Without this, a session
+        // torn down and given a node again inside ONE process would never re-read its durable holds —
+        // the frames would sit in the table, unreleasable, which is indistinguishable from the loss
+        // this unit exists to stop.
+        this.#heldRestored.delete(key);
+        this.#heldReleased.delete(key);
+        this.#diverged.delete(key);
+        this.#counterpartyAddrs.delete(key);
+        this.#redialNotBefore.delete(key);
         this.#highWaterSeq.delete(key);
     }
     /**
@@ -2523,6 +2964,37 @@ export class SessionNodeManager {
      * Called from the SIGTERM / cello logout path (AC-009).
      * SQLite writes complete before this method returns.
      */
+    /**
+     * DOD-M12B-SHUTDOWN-1 — wait for a teardown step, but never forever.
+     *
+     * Every step of shutdown used to be an unbounded `await` on libp2p. That is what makes "the
+     * daemon acknowledged the request but is still running" possible: nothing on the daemon side
+     * emits a word while it hangs, so the operator's own message ("it may be stuck closing sessions
+     * or its database") was a guess. Past the deadline the step is ABANDONED and SAID — the resources
+     * it was closing are reclaimed by the OS on exit, and an exit is worth more than a tidy one.
+     */
+    async #boundedTeardown(work, step, count) {
+        if (count === 0)
+            return;
+        const started = Date.now();
+        let timer;
+        const deadline = new Promise((resolve) => {
+            timer = setTimeout(() => resolve("timeout"), SHUTDOWN_STEP_DEADLINE_MS);
+            timer.unref?.();
+        });
+        const outcome = await Promise.race([work.then(() => "done"), deadline]);
+        if (timer)
+            clearTimeout(timer);
+        if (outcome === "timeout") {
+            this.#logger.error("session.shutdown.step.timeout", {
+                step, count, waitedMs: Date.now() - started,
+                impact: "this teardown step did not finish and was abandoned so the daemon can exit; the OS reclaims what it held",
+            });
+        }
+        else {
+            this.#logger.debug("session.shutdown.step.done", { step, count, tookMs: Date.now() - started });
+        }
+    }
     async gracefulShutdown() {
         // DOD-NAT-REACHABILITY-1: stop the reservation watchdog before anything is torn
         // down — a tick landing mid-shutdown would try to rebuild a receiver we are in
@@ -2533,6 +3005,11 @@ export class SessionNodeManager {
         }
         // Signal any in-flight standing-receiver replacement to self-stop (review M2).
         this.#shuttingDown = true;
+        // DOD-M12B-ACK-1: drop the inbound-content linger resets. The nodes they would reset are being
+        // torn down here anyway, so firing after this point is pure noise on the way out.
+        for (const timer of this.#lingeringStreams)
+            clearTimeout(timer);
+        this.#lingeringStreams.clear();
         // Cancel every armed awaiting-ACK timer so an un-acked send (e.g. a rejected /
         // tampered frame that never produced a `persisted` ACK) does not leave a 20s
         // timer pinning the content + this manager in memory past teardown (review M1).
@@ -2555,7 +3032,9 @@ export class SessionNodeManager {
         else {
             const interruptedAt = new Date(now).toISOString();
             try {
-                this.#db.prepare("UPDATE sessions SET status = 'interrupted', updated_at = ?, interrupted_at = COALESCE(interrupted_at, ?) WHERE status = 'active'").run(now, interruptedAt);
+                this.#db.prepare(
+                // DOD-CAP-SELF-HEAL-1: OURS. Our own shutdown ended these, not the counterparty.
+                "UPDATE sessions SET status = 'interrupted', updated_at = ?, interrupted_at = COALESCE(interrupted_at, ?), interrupted_by = 'local' WHERE status = 'active'").run(now, interruptedAt);
             }
             catch (err) {
                 this.#logger.error("session.interrupt.db.write.failed", {
@@ -2587,14 +3066,22 @@ export class SessionNodeManager {
                 });
             }));
         }
-        await Promise.all(stopPromises);
+        // DOD-M12B-SHUTDOWN-1: BOUNDED. `node.stop()` awaits libp2p's own teardown, which has no
+        // deadline of its own — one connection that will not close holds this, and this holds the whole
+        // daemon. Measured 2026-08-17: `cello logout` acknowledged, then the process was still alive
+        // 30+ seconds later and needed a signal. An abandoned stop costs a socket the OS reclaims when
+        // we exit; an unbounded wait costs the exit itself.
+        await this.#boundedTeardown(Promise.all(stopPromises), "session_nodes", stopPromises.length);
         this.#activeNodes.clear();
         // Evict in-memory per-session caches (trees reload from SQLite; received-content
         // plaintext must not survive shutdown in memory).
         this.#trees.clear();
         this.#receivedContent.clear();
-        // Stop ALL per-agent standing receivers (DOD-LOOP-1).
-        for (const [agentName, sr] of this.#standingReceivers) {
+        // Stop ALL per-agent standing receivers (DOD-LOOP-1). In PARALLEL and BOUNDED: this was a
+        // sequential await per agent with no deadline, so five agents meant five chances for one stuck
+        // libp2p teardown to hold the exit — and it sits between the operator being told the daemon is
+        // stopping and the process actually going.
+        await this.#boundedTeardown(Promise.all([...this.#standingReceivers].map(async ([agentName, sr]) => {
             sr.autoNat.stop();
             try {
                 await sr.node.stop();
@@ -2607,7 +3094,7 @@ export class SessionNodeManager {
                     correlationId: "n/a",
                 });
             }
-        }
+        })), "standing_receivers", this.#standingReceivers.size);
         this.#standingReceivers.clear();
         // Release the SQLite handle so the DB file is no longer held open after shutdown
         // (review L5). Queries guard on `#db === null` and degrade to empty/null.
@@ -2795,7 +3282,21 @@ export class SessionNodeManager {
             // the pre-check above raced (it cannot — DatabaseSync is synchronous), the
             // UPDATE only mutates a row that is still active.
             this.#db
-                .prepare("UPDATE sessions SET status = 'interrupted', updated_at = ?, message_count = ?, interrupted_at = ? WHERE agent_id = ? AND session_id = ? AND status = 'active'")
+                .prepare(
+            // DOD-CAP-SELF-HEAL-1: labelled by SOURCE, because the two are not the same event.
+            //
+            //   relay_frame  — the relay telling us the counterparty went. THEIRS. The D18
+            //                  disconnect-evasion move, and it must keep counting.
+            //   stream_close — OUR witness stream to the relay ended. That fires on a relay restart,
+            //                  a relay fleet roll, or a local network blip. Claiming the counterparty
+            //                  did it means three relay deploys permanently refuse a peer who was
+            //                  never involved — and relay deploys are routine, so it ratchets faster
+            //                  than daemon restarts do.
+            //
+            // `relay_stream_close` is its own label and STILL COUNTS (the bound excuses only 'local'),
+            // because an attacker who can disturb our relay link must not get a free cap reset. It is
+            // recorded honestly rather than blamed on the wrong party.
+            `UPDATE sessions SET status = 'interrupted', updated_at = ?, message_count = ?, interrupted_at = ?, interrupted_by = '${source === "relay_frame" ? "counterparty" : "relay_stream_close"}' WHERE agent_id = ? AND session_id = ? AND status = 'active'`)
                 .run(now, authoritativeCount, interruptedAt, this.#requireAgentId(agentName), sessionId);
         }
         catch (err) {
@@ -3174,6 +3675,9 @@ export class SessionNodeManager {
         for (const addr of addrs) {
             try {
                 await entry.node.dial(addr);
+                // DOD-M12B-REDIAL-1: keep them. They arrived in the signed assignment and were used once
+                // and dropped, which is the reason nothing could ever dial this counterparty again.
+                this.#counterpartyAddrs.set(this.#k(agentName, sessionId), [...addrs]);
                 this.#logger.info("session.transport.connected", {
                     sessionId,
                     addr,
@@ -3246,6 +3750,8 @@ export class SessionNodeManager {
         // the leaf_deliver witness stream / arrival order.
         let orderingS1;
         let orderingS2;
+        // DOD-M12B-INDEX-1: the relay's answer to "where does this message go", carried to the caller.
+        let assignedSeq;
         // DOD-MP-SESSION-RETIRE-1 — the relay's answer SURVIVES to the caller even when the direct send
         // then succeeds. `relay_session_gone` is deliberately not terminal (it also fires for perfectly
         // live sessions whenever the relay restarts, because the relay stores sessions in memory), so
@@ -3263,9 +3769,20 @@ export class SessionNodeManager {
                 if (witnessed.ok) {
                     orderingS1 = witnessed.structure1_cbor;
                     orderingS2 = witnessed.structure2_cbor;
+                    // 1-BASED → 0-BASED. The relay numbers the first leaf of a session 1
+                    // (`relay-node.ts`: `const seq = state.seq_counter + 1`), and this tree is 0-indexed.
+                    // Every RECEIVE path in this file normalises with -1 and says so; the send path took the
+                    // raw number, which puts every comparison against `tree.size()` one position out — so a
+                    // perfectly healthy first message reads as "ahead of the tail" and is held behind a gap
+                    // that does not exist. Do not remove this without changing both receive sites too.
+                    assignedSeq = witnessed.sequence_number - 1;
                     this.#logger.info("session.relay.hash.submitted", {
                         sessionId,
-                        sequenceNumber: witnessed.sequence_number,
+                        // BOTH SPACES, NAMED. The relay's number is 1-based and the leaf index is 0-based, and
+                        // reading one as the other is the defect this milestone exists to stop — so a log that
+                        // carries only "sequenceNumber" invites exactly that mistake on the next investigation.
+                        relaySequence: witnessed.sequence_number,
+                        leafIndex: assignedSeq,
                         correlationId,
                     });
                 }
@@ -3353,8 +3870,14 @@ export class SessionNodeManager {
         // resolves the awaiting timer; on failure (counterparty offline) the hash is already
         // witnessed above, so the caller / TTF path parks the SEALED content to the relay
         // store-and-forward backstop and the recipient recovers it at the witnessed sequence (2b).
+        // Held outside the try so the catch can retire a stream that was opened and then failed to
+        // write. Without it every failure leaks the OUTBOUND half of the stream the receiver-side
+        // `finally` retires — same defect, other end, other cap (64 outbound per protocol per
+        // connection). See the note on #handleContentStream's finally.
+        let sendStream;
         try {
-            const stream = await entry.node.newStream(entry.counterpartySessionPeerId, CELLO_CONTENT_PROTOCOL_ID);
+            const stream = await this.#openContentStream(agentName, sessionId, entry, correlationId);
+            sendStream = stream;
             // AC-001/AC-003: arm the TTF tracking BEFORE the frame goes on the wire. The
             // receiver's `persisted` ACK can come back fast (in-process / low-latency
             // transports), so registering the awaiting entry after send would let the ACK
@@ -3399,9 +3922,17 @@ export class SessionNodeManager {
             // A close that failed for a benign reason costs a redundant park, which the receiver dedups
             // on the content hash. A false delivered costs the message.
             await stream.close();
-            return { ok: true, delivered: true, ...(relayRefusal === undefined ? {} : { relayRefusal }) };
+            this.#clearSessionImpairment(agentName, sessionId, "direct_send", correlationId);
+            return { ok: true, delivered: true, ...(assignedSeq === undefined ? {} : { sequenceNumber: assignedSeq }), ...(relayRefusal === undefined ? {} : { relayRefusal }) };
         }
         catch (err) {
+            this.#markSessionImpaired(agentName, sessionId, { cause: "direct_send", error: err instanceof Error ? err.message : String(err), correlationId });
+            if (sendStream !== undefined) {
+                try {
+                    sendStream.abort(err instanceof Error ? err : new Error(String(err)));
+                }
+                catch { /* already gone */ }
+            }
             // The send failed after (possibly) arming the awaiting tracking — drop it so a
             // never-delivered frame does not later fire a spurious TTF park.
             this.#untrackAwaitingAck(agentName, sessionId, contentHash);
@@ -3412,9 +3943,31 @@ export class SessionNodeManager {
             // can be reported as "dispatched to relay" instead of a raw stream failure — the operator/
             // agent sees the truth (the message IS in flight, just not direct), not a false negative.
             const hashHex = Buffer.from(contentHash).toString("hex");
+            // NAME THE CAUSE. This catch used to discard `err` outright, so a park reported only its exit
+            // point — "dispatched to relay" — and never what went wrong. Measured 2026-08-17: 212 parks on
+            // one daemon, not one of them recording a reason, which is what made a one-way session look
+            // like a protocol mystery for a night.
+            //
+            // `counterpartySessionPeerId` is the load-bearing field. It is recorded ONCE at session
+            // establishment and never refreshed, while a standing receiver is rebuilt with a fresh keypair
+            // on every signaling reconnect — so if the two ever cross, every send goes one-way forever and
+            // nothing says so. With this line that becomes a single grep instead of a night.
+            this.#logger.warn("session.content.direct.send.failed", {
+                agentName,
+                sessionId,
+                contentHash: hashHex,
+                counterpartySessionPeerId: entry.counterpartySessionPeerId,
+                error: err instanceof Error ? err.message : String(err),
+                // "Cannot write to a stream that is closed" names where the write died, never why. The
+                // why is almost always the per-protocol stream cap, and these two numbers are what turn
+                // that from a log-measurement session into a grep.
+                ...this.#streamCensus(entry.node, entry.counterpartySessionPeerId),
+                correlationId,
+            });
             const attempt = await this.#parkContent(agentName, sessionId, hashHex, content, orderingS1, orderingS2);
             if (attempt.outcome === "parked") {
-                return { ok: true, delivered: false, parked: true, ...(relayRefusal === undefined ? {} : { relayRefusal }) };
+                this.#noteImpairmentRetention(agentName, sessionId, "parked");
+                return { ok: true, delivered: false, parked: true, ...(assignedSeq === undefined ? {} : { sequenceNumber: assignedSeq }), ...(relayRefusal === undefined ? {} : { relayRefusal }) };
             }
             // M12-P12: the deposit was refused, and #untrackAwaitingAck above already dropped the
             // in-memory entry — so without this, NOTHING holds the content and the TTF timer that would
@@ -3491,10 +4044,18 @@ export class SessionNodeManager {
             // F3: the two failures are NOT interchangeable to the caller. `reason` is a contract string
             // and stays put; `guidance` carries the difference, because "we are retrying this" and "this
             // message is gone, send it again" demand opposite actions from the operator.
+            //
+            // The same distinction is recorded on the session, because `cello_receive` will be asked
+            // about this later and would otherwise have to guess — and its guess ("it was parked, do not
+            // resend") is the exact opposite of what the lost case needs.
+            this.#noteImpairmentRetention(agentName, sessionId, durable ? "durable" : "lost");
             return {
                 ok: false,
                 reason: "session_stream_unavailable",
                 error: errMsg,
+                // Carried on the failure path too: a DURABLY QUEUED message still owns the position the
+                // relay witnessed for it before delivery was attempted, and its leaf must go there.
+                ...(assignedSeq === undefined ? {} : { sequenceNumber: assignedSeq }),
                 // M12-P13: the machine-readable half of the distinction below. M12-P12 shipped it in the
                 // guidance SENTENCE only, so the callers that have to ACT on it — commit the leaf for a
                 // queued message, never for a lost one — would have had to substring-match English. None
@@ -4129,6 +4690,7 @@ export class SessionNodeManager {
             : this.#witnessedSeq.get(key)?.get(contentHashHex);
         const nextExpected = this.getSessionTree(agentName, sessionId).size();
         if (canonicalSeq !== undefined && canonicalSeq > nextExpected) {
+            this.#ensureHeldRestored(agentName, sessionId);
             let held = this.#heldContent.get(key);
             if (!held) {
                 held = new Map();
@@ -4137,7 +4699,15 @@ export class SessionNodeManager {
             // A terminal block out of canonical order is held WITHOUT delivery (screenedOut): #releaseHeld
             // leafs it at its canonical index when the gap fills, but never buffers it for the agent. This
             // keeps leafIndex === canonicalSeq for screened-out content too (code-review HIGH-1).
-            held.set(canonicalSeq, { content: deliverContent, contentHashHex, correlationId, ...(terminalBlock ? { screenedOut: true } : {}) });
+            // THE PEER'S RAW BYTES RIDE ALONG. Classification (document frame vs conversation) reads
+            // byte 0, and `deliverContent` is the SCREENED copy — for a CBOR frame that is no longer a
+            // map header, so a held document frame was released into the CONVERSATION path: transcript,
+            // doorbell, and `cello_receive` handing an agent raw CBOR as though a person typed it.
+            // The in-order path has always passed these bytes; only the held path dropped them.
+            held.set(canonicalSeq, { content: deliverContent, originalContent: content, contentHashHex, correlationId, ...(terminalBlock ? { screenedOut: true } : {}) });
+            // DOD-M12B-STRAND-1: and to disk, before we answer. The in-memory Map is the working copy;
+            // this row is the one that survives the teardown that used to destroy it.
+            this.#persistHeldContent(agentName, sessionId, canonicalSeq, deliverContent, content, contentHashHex, terminalBlock === true, correlationId);
             this.#logger.info("session.content.held", {
                 sessionId,
                 canonicalSeq,
@@ -4230,6 +4800,9 @@ export class SessionNodeManager {
         if (sequenceNumber < 0)
             return;
         const key = this.#k(agentName, sessionId);
+        // DOD-M12B-SEAL-STUCK-1: this process has now seen this session's ordering state, so an empty
+        // witness map for it means "no gap" rather than "never looked".
+        this.#orderingObserved.add(key);
         let map = this.#witnessedSeq.get(key);
         if (!map) {
             map = new Map();
@@ -4286,6 +4859,16 @@ export class SessionNodeManager {
      */
     sealReadiness(agentName, sessionId) {
         const key = this.#k(agentName, sessionId);
+        // DOD-M12B-STRAND-1: hydrate first. An under-counted `heldCount` reports a gapped session as
+        // READY, and this gate's whole purpose is to stop a short chain being signed — the counterparty
+        // answers `leaf_count_mismatch`, which is TERMINAL and costs the receipt permanently. Failing
+        // open here is the one outcome worse than refusing a healthy close.
+        //
+        // Hydrate WITHOUT releasing: this is a read path now — the status surface asks it for every
+        // active session — and a read that appends leaves, advances the root and rings the doorbell is
+        // a diagnostic command that delivers messages. The release still runs on every path that was
+        // going to mutate anyway.
+        this.#ensureHeldRestored(agentName, sessionId, { release: false });
         const treeSize = this.getSessionTree(agentName, sessionId).size();
         const highWaterSeq = this.#highWaterSeq.get(key) ?? -1;
         const heldCount = this.#heldContent.get(key)?.size ?? 0;
@@ -4305,7 +4888,107 @@ export class SessionNodeManager {
         // ordering authority has committed and this tree has not — no arithmetic, no space mismatch,
         // and it cannot go negative.
         const missingLeaves = this.#witnessedSeq.get(key)?.size ?? 0;
-        return { ready: missingLeaves === 0 && heldCount === 0, treeSize, highWaterSeq, heldCount, missingLeaves };
+        // DOD-M12B-INDEX-1: our OWN held sends are counted separately. They block a seal just as a
+        // received hold does, but they are not "a message from the counterparty that has not arrived" —
+        // and a refusal that calls them that tells the operator to wait for something already in hand,
+        // which is how a close-retry loop ends at force-abandon and no receipt.
+        let heldOwn = 0;
+        for (const e of this.#heldContent.get(key)?.values() ?? [])
+            if (e.origin === "sent")
+                heldOwn++;
+        return {
+            ready: missingLeaves === 0 && heldCount === 0,
+            treeSize, highWaterSeq, heldCount, missingLeaves,
+            heldOwn, heldReceived: heldCount - heldOwn,
+        };
+    }
+    /**
+     * DOD-M12B-SEAL-STUCK-1 — the operator-facing answer to "can this session be closed?".
+     *
+     * THREE STATES, because there are three answers. `sealReadiness` above returns a boolean plus raw
+     * counters, and both of its counters are easy to misread on a surface:
+     *
+     *  - `missingLeaves` is `#witnessedSeq.size`, which is every position the relay witnessed that
+     *    this tree has not appended — and a HELD frame keeps its witness entry. So it INCLUDES the
+     *    held ones. Reporting it beside `heldCount` counts the same message twice and labels one copy
+     *    "never received" when it is sitting on our own disk. Split here into what each actually is.
+     *  - Neither counter survives a restart on its own: `#witnessedSeq` is memory-only. Held content
+     *    is durable since DOD-M12B-STRAND-1, but a position the relay witnessed for content that
+     *    never arrived leaves no trace. So for a session carrying leaves this process did not watch
+     *    arrive, "clean" is unknowable — and saying `ready` there invites a close that gets
+     *    `leaf_count_mismatch` back, which is terminal and costs the receipt for good.
+     */
+    sealReadinessView(agentName, sessionId) {
+        const key = this.#k(agentName, sessionId);
+        const r = this.sealReadiness(agentName, sessionId);
+        if (!r.ready) {
+            const oldestHeldMs = this.#oldestHeldMs(agentName, sessionId);
+            return {
+                state: "blocked",
+                // The witness map counts a held frame until it is appended, so subtract the RECEIVED holds
+                // to avoid reporting one message twice. NOT the own-sends: the witness map only ever
+                // carries counterparty leaves, so subtracting ours would push this count below the truth.
+                awaitingArrival: Math.max(0, r.missingLeaves - r.heldReceived),
+                heldBehindGap: r.heldCount,
+                oldestHeldMs,
+            };
+        }
+        if (this.#diverged.has(key)) {
+            // NOT `ready`. The tree is ahead of the relay's counter for good, so a close here signs a root
+            // the counterparty answers `leaf_count_mismatch` to — terminal, and the receipt is gone. The
+            // raw counters cannot see this: nothing is missing and nothing is held.
+            return { state: "unknown", reason: "record_diverged_from_relay" };
+        }
+        if (r.treeSize > 0 && !this.#orderingObserved.has(key)) {
+            return {
+                state: "unknown",
+                reason: "witness_state_predates_daemon_start",
+            };
+        }
+        return { state: "ready" };
+    }
+    /** DOD-M12B-INDEX-1 — this agent's own K_local pubkey, for attributing its own held content.
+     *  Null when it cannot be resolved: an UNATTRIBUTED annex row is true, a falsely attributed one
+     *  is not, and this is the record that outlives the session. */
+    #ownPubkeyHex(agentName) {
+        if (!this.#db)
+            return null;
+        try {
+            // BY agent_id, never by agent_name. The name is a mutable, reuse-freed display label, and
+            // scoping on it hands one identity's rows to another keypair (DOD-AGENT-ID-JOINKEY-1).
+            const row = this.#db
+                .prepare("SELECT k_local_pubkey FROM agents WHERE agent_id = ?")
+                .get(this.#requireAgentId(agentName));
+            return row?.k_local_pubkey ?? null;
+        }
+        catch (err) {
+            // An unattributed annex row is truthful; an unattributed row nobody knows about is not. This
+            // throws for a retired agent, and without a line here EVERY own held message would land in
+            // the record that outlives the session with no sender and no explanation.
+            this.#logger.warn("session.own_pubkey.unresolved", {
+                agentName,
+                error: err instanceof Error ? err.message : String(err),
+                impact: "this agent's own held content will be annexed without a sender",
+            });
+            return null;
+        }
+    }
+    /** DOD-M12B-SEAL-STUCK-1 — how long the oldest held frame for this session has been waiting, or
+     *  null when nothing is held. This is what separates "stuck since this morning" from "in flight
+     *  40 ms ago", and without it a healthy mid-conversation window reads as a stranded session. */
+    #oldestHeldMs(agentName, sessionId) {
+        if (!this.#db)
+            return null;
+        try {
+            const row = this.#db.prepare("SELECT MIN(held_at) AS oldest FROM held_content WHERE agent_id = ? AND session_id = ?").get(this.#requireAgentId(agentName), sessionId);
+            if (!row || row.oldest === null)
+                return null;
+            return Date.now() - row.oldest;
+        }
+        catch {
+            // A diagnostic detail must not be able to break the surface it decorates.
+            return null;
+        }
     }
     /** DOD-MSG-4 / DAEMON-004: append a verified message leaf and buffer it for cello_receive. */
     #appendVerifiedContent(agentName, sessionId, content, contentHashHex, senderPubkey, correlationId, 
@@ -4429,6 +5112,520 @@ export class SessionNodeManager {
         }
         return { leafIndex };
     }
+    /**
+     * DOD-M12B-ABANDON-NOTIFY-1 — tell the counterparty we have hung up. Best effort, never blocking.
+     *
+     * A force-abandon marks the session terminal HERE and did nothing else, so the other side kept
+     * its half live, kept retrying delivery into it, and kept trying to re-establish — forever,
+     * because nothing would ever answer. That is what produced the 2026-08-17 notification storm:
+     * surviving halves calling continuously while the operator saw connection requests from agents
+     * nobody was driving.
+     *
+     * BEST EFFORT, and every caller must treat it that way. A peer that is offline cannot be told, so
+     * this is an improvement on silence rather than a guarantee — and it must never delay or fail the
+     * abandon, which is the operator's escape hatch out of a session that can never seal.
+     */
+    async notifyCounterpartyAbandon(agentName, sessionId, correlationId) {
+        const entry = this.#activeNodes.get(this.#k(agentName, sessionId));
+        if (!entry) {
+            // NAMES ITS CAUSE, and it is not the network. An `interrupted` session has no node — the
+            // restart sweep and markInterrupted both tear it down — and `interrupted` is exactly the
+            // status force-abandon exists for. Reporting this as "could not be reached" sends the
+            // operator to debug a connection when the answer is in our own process. At INFO, not debug,
+            // because it is the common case and it changes what the operator is told.
+            this.#logger.info("session.abandon.notice.skipped", {
+                agentName, sessionId, reason: "no_local_node", correlationId,
+                impact: "this side had already torn the session down, so there was nothing to send on — the counterparty was not told",
+            });
+            return { told: false, reason: "no_local_node" };
+        }
+        let stream;
+        try {
+            // Through the RE-DIAL path, not a bare newStream. A session worth force-abandoning is very
+            // often one whose connection blipped — the peer is online and calling us, which is the whole
+            // complaint — so one demand-driven dial is the difference between telling them and not.
+            stream = await this.#openContentStream(agentName, sessionId, entry, correlationId);
+            // Typed against protocol-types so the shape cannot drift from the declaration the receiving
+            // side (and any second client implementation) reads.
+            const notice = {
+                type: "session_abandoned_notice",
+                session_id: sessionId,
+                ...(correlationId === undefined ? {} : { correlation_id: correlationId }),
+            };
+            const frame = encodeCbor(notice);
+            stream.send(lp.encode.single(frame));
+            await stream.close();
+            this.#logger.info("session.abandon.notice.sent", { agentName, sessionId, correlationId });
+            return { told: true, reason: "sent" };
+        }
+        catch (err) {
+            if (stream !== undefined) {
+                try {
+                    stream.abort(err instanceof Error ? err : new Error(String(err)));
+                }
+                catch { /* already gone */ }
+            }
+            this.#logger.warn("session.abandon.notice.failed", {
+                agentName, sessionId, correlationId,
+                error: err instanceof Error ? err.message : String(err),
+                impact: "the counterparty was not told and may keep calling until it gives up",
+            });
+            return { told: false, reason: "send_failed" };
+        }
+    }
+    /**
+     * DOD-M12B-ABANDON-NOTIFY-1 — the receiving half: our counterparty has abandoned, so retire.
+     *
+     * RETIRING IS NOT DELETING. The counterparty walking away forfeits the notarized receipt; it must
+     * not also cost the operator the record of what was actually said. The transcript and the tree
+     * stay exactly as they are.
+     *
+     * Only an `active` or `interrupted` session moves. A SEALED session has a notarized receipt and
+     * must never be turned into an abandoned one by a late or duplicated notice — that would destroy
+     * the artifact this protocol exists to produce. An unknown session is refused rather than
+     * created: an authenticated stream proves who is speaking, not that a session exists.
+     */
+    async retireOnCounterpartyAbandon(agentName, sessionId, correlationId) {
+        const record = this.getSessionRecord(agentName, sessionId);
+        if (!record) {
+            this.#logger.warn("session.abandon.notice.unknown_session", { agentName, sessionId, correlationId });
+            return false;
+        }
+        if (record.status !== "active" && record.status !== "interrupted") {
+            this.#logger.debug("session.abandon.notice.ignored", {
+                agentName, sessionId, status: record.status, correlationId,
+                reason: "session already terminal",
+            });
+            return false;
+        }
+        // THE TRANSPORT IS RETIRED. THE SESSION IS NOT.
+        //
+        // The first build flipped the status to `abandoned`, and that was wrong twice over. It handed
+        // the abandoning party a button that DENIES US OUR RECEIPT: the unilateral seal exists for
+        // exactly this case — "the counterparty never co-closes" — and produces a notarized certificate
+        // after a grace period, but `cello_close_session` refuses an `abandoned` session outright. So
+        // one frame from them destroyed a recovery path that already existed, remotely and for free.
+        // Today the abandoner can only go silent, and going silent is what the unilateral seal was
+        // built to survive.
+        //
+        // What the DoD actually asks for is that we stop calling them. That is a transport concern:
+        // mark it, stop re-dialling, stop retrying delivery — and leave the session sealable.
+        const marked = this.#markCounterpartyAbandoned(agentName, sessionId);
+        if (!marked)
+            return false;
+        // The addresses go, so the demand-driven re-dial has nothing to dial. This is the storm.
+        const key = this.#k(agentName, sessionId);
+        this.#counterpartyAddrs.delete(key);
+        this.#logger.warn("session.counterparty.abandoned", {
+            agentName, sessionId, priorStatus: record.status, correlationId,
+            impact: "the counterparty ended this session on their side, so nothing more will arrive and replies cannot reach them — this side stops calling. The session is NOT terminal: a unilateral seal is still available, and the transcript is intact",
+        });
+        // AWAITED, and `retireSessionNode` NOT `destroySessionNode`. The latter writes the status back
+        // — `error` maps to `interrupted` — a few hundred milliseconds later, which silently undid the
+        // whole unit; the former is the method that tears a node down without touching the status, and
+        // it is what the local force-abandon path already uses.
+        await this.retireSessionNode(agentName, sessionId);
+        return true;
+    }
+    /** DOD-M12B-ABANDON-NOTIFY-1 — durable "they hung up" marker. Not a status: the session stays
+     *  sealable, which is the whole point of not making this terminal. */
+    #markCounterpartyAbandoned(agentName, sessionId) {
+        if (!this.#db)
+            return false;
+        try {
+            const res = this.#db
+                .prepare("UPDATE sessions SET counterparty_abandoned_at = ?, updated_at = ? WHERE agent_id = ? AND session_id = ? AND counterparty_abandoned_at IS NULL")
+                .run(Date.now(), Date.now(), this.#requireAgentId(agentName), sessionId);
+            // No rows changed means it was already marked — a duplicated notice, which must not
+            // re-announce. "Did not throw" is not "landed"; the row count is the answer.
+            return Number(res?.changes ?? 0) > 0;
+        }
+        catch (err) {
+            this.#logger.error("session.counterparty.abandoned.write.failed", {
+                agentName, sessionId,
+                error: err instanceof Error ? err.message : String(err),
+                impact: "this side will go on trying to reach a counterparty that has hung up",
+            });
+            return false;
+        }
+    }
+    /**
+     * DOD-CAP-SELF-HEAL-1 — the numbers behind a cap refusal, for the OPERATOR'S alarm only.
+     *
+     * Kept off `checkUnknownSenderAcceptanceBound`'s return on purpose. That refusal is byte-identical
+     * across tiers by design (DOD-TIER-3) so a blocked party cannot tell blocking from throttling;
+     * attaching the counts to it would put the oracle straight into the value the refusal path
+     * carries. This is a separate, purely local read, and nothing it returns crosses the wire.
+     */
+    capDiagnostics(agentName, counterpartyPubkey) {
+        const tier = this.getTier(agentName, counterpartyPubkey);
+        const cap = this.resolveTierBound(agentName, tier, "max_sessions");
+        const counted = this.countActiveSessionsForCounterparty(agentName, counterpartyPubkey);
+        return {
+            tier, cap, counted,
+            // How many to close to get UNDER the cap — not how many exist. At 5 against a cap of 3 the
+            // answer is 3, and "close 5" tells the operator to do more than the job needs.
+            mustClear: Math.max(0, counted - cap + 1),
+            blocked: tier === TIER.BLOCKED,
+        };
+    }
+    /** DOD-CAP-SELF-HEAL-1: the sessions with this counterparty that are consuming cap slots, oldest
+     *  first. The operator is told to close some — this is WHICH, because "close three of them" with
+     *  no list is not an instruction they can follow. */
+    sessionsConsumingCap(agentName, counterpartyPubkey, limit = 10) {
+        if (!this.#db)
+            return [];
+        try {
+            const rows = this.#db.prepare(`SELECT session_id FROM sessions WHERE agent_id = ? AND counterparty_pubkey = ? AND ${CAP_COUNTS()}
+         ORDER BY updated_at ASC LIMIT ?`).all(this.#requireAgentId(agentName), counterpartyPubkey, limit);
+            return rows.map((r) => r.session_id);
+        }
+        catch {
+            return [];
+        }
+    }
+    /** DOD-M12B-ABANDON-NOTIFY-1: has the counterparty told us they hung up? */
+    counterpartyAbandonedAt(agentName, sessionId) {
+        if (!this.#db)
+            return null;
+        try {
+            const row = this.#db
+                .prepare("SELECT counterparty_abandoned_at FROM sessions WHERE agent_id = ? AND session_id = ?")
+                .get(this.#requireAgentId(agentName), sessionId);
+            return row?.counterparty_abandoned_at ?? null;
+        }
+        catch {
+            return null;
+        }
+    }
+    /**
+     * DOD-M12B-REDIAL-1 — open a content stream, re-dialling once if the connection has gone.
+     *
+     * `newStream` never dials. It looks for an already-open connection filed under the recorded peer
+     * id and throws `connection_lost` when there is none — and NOTHING re-dialled: not on
+     * `session.liveness.changed → gone`, not on signaling reconnect, not on agent offline→online, not
+     * in the drain hook. So one blip and that session parked EVERY message for the rest of its life,
+     * on both sides, permanently. The relay backstop kept the messages moving, which is exactly why
+     * it hid: nothing was lost, the conversation just stopped being a conversation.
+     *
+     * DEMAND-DRIVEN, never a timer. A background re-dial loop is what produced the 2026-08-17
+     * notification storm — surviving halves of abandoned sessions dialling continuously while the
+     * operator saw connection requests from agents nobody was driving. This fires only when a send
+     * actually needs the connection, and a cooldown bounds a burst against a peer that is genuinely
+     * gone: five sends cost one dial, not five.
+     */
+    async #openContentStream(agentName, sessionId, entry, correlationId) {
+        const attempt = async () => {
+            if (this.#connectionLossRemaining > 0) {
+                this.#connectionLossRemaining -= 1;
+                throw { reason: "no_connection", message: "injected connection loss" };
+            }
+            return entry.node.newStream(entry.counterpartySessionPeerId, CELLO_CONTENT_PROTOCOL_ID);
+        };
+        try {
+            return await attempt();
+        }
+        catch (err) {
+            const reason = err?.reason;
+            // ONLY for a missing connection, and `no_connection` is the reason that means exactly that.
+            // NOT `connection_lost`, which is the transport's catch-all default and therefore also covers
+            // a stream that failed on a healthy connection — the per-protocol stream cap of
+            // DOD-M12B-ACK-1. Dialling there fixes nothing and shows the counterparty a connection
+            // request caused by a defect on this side, which is the storm this unit exists to avoid.
+            if (reason !== "no_connection")
+                throw err;
+            const key = this.#k(agentName, sessionId);
+            const addrs = this.#counterpartyAddrs.get(key);
+            if (!addrs || addrs.length === 0) {
+                // ABSENT IS NOT FINE, and it is not silent. A session we never dialled — the responder's
+                // half — has no addresses to dial back with, and that is a real limitation the operator
+                // should be able to see rather than infer from a park.
+                this.#logger.warn("session.transport.redial.unavailable", {
+                    sessionId, agentName, correlationId,
+                    impact: "the direct path is down and this side holds no address for the counterparty, so every send parks until they re-establish",
+                });
+                throw err;
+            }
+            const now = Date.now();
+            const notBefore = this.#redialNotBefore.get(key) ?? 0;
+            if (now < notBefore) {
+                this.#logger.debug("session.transport.redial.cooldown", {
+                    sessionId, agentName, retryInMs: notBefore - now, correlationId,
+                });
+                throw err;
+            }
+            this.#redialNotBefore.set(key, now + REDIAL_COOLDOWN_MS);
+            this.#logger.info("session.transport.redial.attempted", { sessionId, agentName, addrs: addrs.length, correlationId });
+            const reconnected = await this.connectToCounterparty(agentName, sessionId, addrs);
+            if (!reconnected.ok) {
+                this.#logger.warn("session.transport.redial.failed", {
+                    sessionId, agentName, reason: reconnected.reason, error: reconnected.error, correlationId,
+                });
+                throw err;
+            }
+            this.#logger.info("session.transport.redial.succeeded", { sessionId, agentName, correlationId });
+            // Cleared so the NEXT blip is repaired immediately: the cooldown exists to bound a dead peer,
+            // not to make a live one wait.
+            this.#redialNotBefore.delete(key);
+            return attempt();
+        }
+    }
+    /**
+     * DOD-M12B-INDEX-1 — commit THIS agent's own leaf at the position the relay assigned it.
+     *
+     * The receiver has always enforced "leaf index === canonical position": content witnessed ahead
+     * of the next expected leaf is held, not appended out of order. The sender never did. It had the
+     * position in hand — the relay answers about 4 ms before the append — and called a push-only
+     * append that puts the leaf at the tail whatever the tail happens to be. While its own tree has
+     * no gap the two agree and nothing shows; the first gap puts its leaf at someone else's index,
+     * parts its root from the counterparty's, and the next seal gets `leaf_count_mismatch`, which is
+     * terminal.
+     *
+     * DELIVERY IS NOT DEFERRED BY THIS. The caller has already put the bytes on the wire; only the
+     * leaf waits for its slot, exactly as a received message does. Holding our own send is only
+     * affordable because holds are durable (DOD-M12B-STRAND-1) — before that it would have risked
+     * losing the message outright.
+     *
+     * `assignedSeq` absent means no ordering authority answered. That is the documented degradation
+     * and it appends in arrival order as before: with no position there is no discipline to enforce,
+     * and refusing would take messaging down whenever the relay is unreachable.
+     */
+    placeOwnLeaf(agentName, sessionId, contentHashHex, sentBytes, assignedSeq, correlationId, kind = "msg") {
+        // Hydrate before reading the frontier: a durable hold this process has not read back yet would
+        // make the tree look further along than it is.
+        this.#ensureHeldRestored(agentName, sessionId);
+        const nextExpected = this.getSessionTree(agentName, sessionId).size();
+        if (assignedSeq === undefined) {
+            const { leafIndex } = this.appendSessionLeaf(agentName, sessionId, kind, contentHashHex, correlationId);
+            return { placed: true, leafIndex };
+        }
+        if (assignedSeq === nextExpected) {
+            const { leafIndex } = this.appendSessionLeaf(agentName, sessionId, kind, contentHashHex, correlationId);
+            return { placed: true, leafIndex };
+        }
+        if (assignedSeq < nextExpected) {
+            // THE TREE AND THE RELAY HAVE ALREADY DIVERGED, and refusing here does not undo that.
+            //
+            // This side is AHEAD of the relay's counter, which happens by design: a message whose relay
+            // submit failed still appends unwitnessed (the documented degradation). From then on every
+            // ack comes back behind our frontier and the two can never agree again — the seal was already
+            // lost at the unwitnessed append, not here.
+            //
+            // So the choice is between a record that is short by every subsequent message and one that is
+            // complete but skewed. Appending at the tail keeps the operator's own words in their own
+            // transcript, which is worth more than a tidiness the roots cannot recover anyway; writing
+            // over the assigned slot is the one thing never done, because that rewrites a leaf a root has
+            // already been computed over. The divergence is reported at ERROR and carried to the caller
+            // rather than dressed up as an ordinary success.
+            this.#logger.error("session.tree.position_behind_frontier", {
+                agentName, sessionId, assignedSeq, nextExpected, contentHash: contentHashHex, correlationId,
+                impact: "this side's tree is ahead of the relay's counter, so the two can no longer agree on a root — the message is kept in the local record and this session can no longer be sealed bilaterally",
+            });
+            this.#diverged.add(this.#k(agentName, sessionId));
+            const { leafIndex } = this.appendSessionLeaf(agentName, sessionId, kind, contentHashHex, correlationId);
+            return { placed: true, leafIndex, diverged: true };
+        }
+        // Ahead of the tail: hold it, exactly as the receiver holds theirs, and let #releaseHeld put it
+        // in at its own index when the gap fills.
+        const key = this.#k(agentName, sessionId);
+        let held = this.#heldContent.get(key);
+        if (!held) {
+            held = new Map();
+            this.#heldContent.set(key, held);
+        }
+        held.set(assignedSeq, { content: sentBytes, contentHashHex, correlationId, origin: "sent", kind });
+        this.#persistHeldContent(agentName, sessionId, assignedSeq, sentBytes, sentBytes, contentHashHex, false, correlationId, "sent", kind);
+        this.#logger.info("session.content.held", {
+            sessionId, canonicalSeq: assignedSeq, nextExpected, gap: assignedSeq - nextExpected,
+            origin: "sent", correlationId,
+        });
+        return { placed: false, heldAt: assignedSeq };
+    }
+    /**
+     * DOD-M12B-STRAND-1 — write one held frame to the durable store.
+     *
+     * LOGS LOUD, does not refuse. The caller answers `held: true` either way, and that is correct:
+     * held content is never `persisted`-acked, so the sender keeps its copy and retries whether or
+     * not this row lands. What a failure costs is the restart case — the frame is memory-only again,
+     * exactly as it was before this unit — so it is reported at ERROR here and counted again by the
+     * teardown alarm, and never allowed to look like a success.
+     */
+    #persistHeldContent(agentName, sessionId, canonicalSeq, deliverContent, originalContent, contentHashHex, screenedOut, correlationId, origin = "received", leafKind = "msg") {
+        if (!this.#db)
+            return;
+        try {
+            // A position may legitimately be re-written by a redelivery of the SAME frame. Different
+            // content at the same relay position means the relay contradicted itself, and destroying the
+            // first copy silently is not an option for verified content.
+            const existing = this.#db.prepare("SELECT content_hash_hex FROM held_content WHERE agent_id = ? AND session_id = ? AND canonical_seq = ?").get(this.#requireAgentId(agentName), sessionId, canonicalSeq);
+            if (existing && existing.content_hash_hex !== contentHashHex) {
+                this.#logger.error("session.content.held.position_conflict", {
+                    agentName, sessionId, canonicalSeq, correlationId,
+                    existingContentHash: existing.content_hash_hex, incomingContentHash: contentHashHex,
+                    impact: "two different frames claim one canonical position — the earlier held copy is being replaced",
+                });
+            }
+            this.#db.prepare(`INSERT OR REPLACE INTO held_content
+           (agent_id, session_id, canonical_seq, content_blob, original_blob, content_hash_hex, screened_out, correlation_id, held_at, origin, leaf_kind)
+         VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)`).run(this.#requireAgentId(agentName), sessionId, canonicalSeq, Buffer.from(deliverContent), Buffer.from(originalContent), contentHashHex, screenedOut ? 1 : 0, correlationId ?? null, Date.now(), origin, leafKind);
+        }
+        catch (err) {
+            this.#logger.error("session.content.held.persist.failed", {
+                agentName, sessionId, canonicalSeq, contentHash: contentHashHex, correlationId,
+                impact: "this frame is held IN MEMORY ONLY and will be destroyed if the daemon restarts",
+                error: err instanceof Error ? err.message : String(err),
+            });
+        }
+    }
+    /** DOD-M12B-STRAND-1 — drop one held frame from the durable store, on release or on a refusal
+     *  that supersedes it. A row that outlives its release re-appends on the next boot. */
+    #deleteHeldContent(agentName, sessionId, canonicalSeq) {
+        if (!this.#db)
+            return;
+        try {
+            this.#db.prepare("DELETE FROM held_content WHERE agent_id = ? AND session_id = ? AND canonical_seq = ?").run(this.#requireAgentId(agentName), sessionId, canonicalSeq);
+        }
+        catch (err) {
+            this.#logger.error("session.content.held.delete.failed", {
+                agentName, sessionId, canonicalSeq,
+                impact: "the released frame's durable row survives and will be re-appended on the next boot",
+                error: err instanceof Error ? err.message : String(err),
+            });
+        }
+    }
+    /**
+     * DOD-M12B-STRAND-1 — restore this session's held frames into memory.
+     *
+     * Called when a session node is (re)created, which is the moment the session becomes able to
+     * append again. Loading at daemon boot instead would be wrong for the same reason the old code
+     * was wrong: a frame is only releasable against a tree, and the tree is loaded per session.
+     *
+     * A frame whose content the tree ALREADY HOLDS at that position is dropped — re-appending it
+     * would change the root a seal signs over. That test is `hashAt`, not `canonical_seq < frontier`:
+     * the two counters are different spaces and this file documents them drifting, so the index
+     * comparison alone would destroy a frame the tree never held. See the drift branch below.
+     */
+    #ensureHeldRestored(agentName, sessionId, opts) {
+        const key = this.#k(agentName, sessionId);
+        if (!this.#heldRestored.has(key)) {
+            // Set BEFORE restoring, so the #ensureHeldRestored inside #releaseHeld below returns straight
+            // away instead of recursing.
+            this.#heldRestored.add(key);
+            this.#restoreHeldContent(agentName, sessionId);
+        }
+        // A RESTORED FRAME MAY ALREADY BE IN ORDER, and nothing else would ever notice.
+        //
+        // #releaseHeld has one caller: the tail of a successful inbound ingest. Every other way the tree
+        // grows — an outbound send leaf, a queued or rejected leaf — advances the frontier without
+        // draining. While holds died with the session node that cost seconds; now the hold is durable,
+        // so the stall is durable too: the counterparty's message sits on disk at exactly the next slot,
+        // is never delivered, and `sealReadiness` counts it, so the session cannot close either.
+        // Undeliverable AND unsealable, forever, from one restart.
+        //
+        // TRACKED SEPARATELY FROM THE HYDRATION. A read-only caller (the status surface) hydrates and
+        // must NOT release — otherwise `cello status` appends leaves, advances the session root, writes
+        // transcript rows and rings the doorbell, which makes a diagnostic command the thing that
+        // delivers messages. One shared flag would also let that read CONSUME the release the next real
+        // ingest was going to perform, which is the stall above, reintroduced.
+        if (opts?.release === false)
+            return;
+        if (this.#heldReleased.has(key))
+            return;
+        this.#heldReleased.add(key);
+        if (this.#heldContent.get(key)?.size) {
+            const counterparty = this.getSessionRecord(agentName, sessionId)?.counterparty_pubkey;
+            if (counterparty)
+                this.#releaseHeld(agentName, sessionId, counterparty);
+        }
+    }
+    #restoreHeldContent(agentName, sessionId) {
+        if (!this.#db)
+            return;
+        let rows;
+        try {
+            rows = this.#db.prepare(`SELECT canonical_seq, content_blob, original_blob, content_hash_hex, screened_out, correlation_id, origin, leaf_kind
+           FROM held_content WHERE agent_id = ? AND session_id = ? ORDER BY canonical_seq ASC`).all(this.#requireAgentId(agentName), sessionId);
+        }
+        catch (err) {
+            this.#logger.error("session.content.held.restore.failed", {
+                agentName, sessionId,
+                impact: "verified content held before the restart is not in memory and cannot be released",
+                error: err instanceof Error ? err.message : String(err),
+            });
+            return;
+        }
+        if (rows.length === 0)
+            return;
+        const key = this.#k(agentName, sessionId);
+        let held = this.#heldContent.get(key);
+        if (!held) {
+            held = new Map();
+            this.#heldContent.set(key, held);
+        }
+        const tree = this.getSessionTree(agentName, sessionId);
+        const frontier = tree.size();
+        let restored = 0;
+        let superseded = 0;
+        let drifted = 0;
+        for (const row of rows) {
+            // ASK THE TREE WHAT IS AT THAT POSITION — do not infer it from the index.
+            //
+            // `canonical_seq` is the RELAY's sequence space; `frontier` is this tree's msg-leaf count.
+            // The two drift, on purpose and by documented cases: the relay counts CTRL leaves the tree
+            // never appends, and a first message whose relay submit failed leaves the tree one ahead.
+            // Under drift `canonical_seq < frontier` is TRUE for a frame the tree has never held, and
+            // deleting on that comparison destroys verified content while reporting it as tidy-up —
+            // the exact failure this unit exists to end, reintroduced on the recovery path.
+            const occupant = tree.hashAt(row.canonical_seq);
+            if (occupant === row.content_hash_hex) {
+                this.#deleteHeldContent(agentName, sessionId, row.canonical_seq);
+                superseded++;
+                continue;
+            }
+            if (row.canonical_seq < frontier) {
+                // The position is taken by DIFFERENT content. The frame cannot be appended (that would
+                // rewrite a committed leaf) and must not be deleted (it is verified content nobody else
+                // holds), so it goes to the annex that exists for exactly this — content that arrived for
+                // a chain that can no longer carry it — and only then does the row go.
+                const annexed = this.recordSealedAnnex(agentName, sessionId, row.content_hash_hex, new Uint8Array(row.content_blob), 
+                // DOD-M12B-INDEX-1: our own held send is attributed to US, never to the counterparty.
+                row.origin === "sent"
+                    ? this.#ownPubkeyHex(agentName)
+                    : this.getSessionRecord(agentName, sessionId)?.counterparty_pubkey ?? null);
+                this.#logger.error("session.content.held.position_drifted", {
+                    agentName, sessionId, canonicalSeq: row.canonical_seq, frontier,
+                    contentHash: row.content_hash_hex, occupant, annexed,
+                    correlationId: row.correlation_id ?? undefined,
+                    impact: annexed
+                        ? "the relay's position for this frame is occupied by different content — it cannot join the chain and is readable only from the annex"
+                        : "the relay's position for this frame is occupied by different content AND the annex write failed — the durable row is kept rather than destroyed",
+                });
+                if (annexed)
+                    this.#deleteHeldContent(agentName, sessionId, row.canonical_seq);
+                drifted++;
+                continue;
+            }
+            held.set(row.canonical_seq, {
+                content: new Uint8Array(row.content_blob),
+                ...(row.original_blob ? { originalContent: new Uint8Array(row.original_blob) } : {}),
+                contentHashHex: row.content_hash_hex,
+                ...(row.correlation_id ? { correlationId: row.correlation_id } : {}),
+                ...(row.screened_out ? { screenedOut: true } : {}),
+                ...(row.origin === "sent" ? { origin: "sent" } : {}),
+                ...(row.leaf_kind === "doc" ? { kind: "doc" } : {}),
+            });
+            restored++;
+        }
+        if (held.size === 0)
+            this.#heldContent.delete(key);
+        this.#logger.info("session.content.held.restored", {
+            agentName, sessionId, restored, superseded, drifted, frontier,
+            canonicalSeqs: [...held.keys()].sort((a, b) => a - b),
+            // The flow ids of the frames that came back, so a restored message ties to the
+            // `session.content.held` that opened its flow before the restart.
+            correlationIds: rows.map((r) => r.correlation_id).filter((c) => c !== null),
+        });
+    }
     /**
      * DOD-MSG-4: drain held out-of-order content in canonical order. After a leaf is appended, any
      * held entry whose canonical sequence equals the new next-expected index is now in order — append
@@ -4436,6 +5633,10 @@ export class SessionNodeManager {
      */
     #releaseHeld(agentName, sessionId, senderPubkey) {
         const key = this.#k(agentName, sessionId);
+        // DOD-M12B-STRAND-1: hydrate before scanning. Restoring eagerly at session-node creation put
+        // it behind writes that can fail — one failed `sessions` row upsert and the frames stayed on
+        // disk, invisible, which is the same outcome as losing them.
+        this.#ensureHeldRestored(agentName, sessionId);
         const held = this.#heldContent.get(key);
         if (!held)
             return 0;
@@ -4446,13 +5647,47 @@ export class SessionNodeManager {
             if (!entry)
                 break;
             held.delete(nextExpected);
-            // A screened-out (terminal-blocked) held entry leafs at its canonical index but is NEVER
-            // buffered for the agent; a normal held entry buffers + leafs (code-review HIGH-1).
-            if (entry.screenedOut) {
+            // DOD-M12B-STRAND-1: released content leaves the durable store in the same breath. A row
+            // that outlives its release would re-append the same content on the next boot — growing the
+            // tree and changing a root that has already been signed.
+            this.#deleteHeldContent(agentName, sessionId, nextExpected);
+            // DOD-M12B-INDEX-1: OUR OWN held message. It leafs at its canonical index and is transcribed
+            // as SENT — never routed down the received path, which would attribute our words to the
+            // counterparty in the sealed record and hand them back to our own agent as inbound.
+            if (entry.origin === "sent") {
+                // The KIND the leaf was placed with, not a hardcoded "msg" — a document leaf that had to
+                // wait for its position must come back as a document leaf, or the two sides disagree about
+                // what the chain contains.
+                this.appendSessionLeaf(agentName, sessionId, entry.kind ?? "msg", entry.contentHashHex, entry.correlationId);
+                // A DOCUMENT frame takes a leaf and NO transcript row — matching what the immediate-append
+                // path does for one. Writing one would put raw CBOR into the operator's transcript as
+                // something they said, which is the same attribution failure as releasing it inbound.
+                if (entry.kind === "doc") {
+                    released++;
+                    this.#logger.info("session.content.released", {
+                        sessionId, sequenceNumber: nextExpected, leafKind: "doc", correlationId: entry.correlationId,
+                    });
+                    if (held.size === 0) {
+                        this.#heldContent.delete(key);
+                        break;
+                    }
+                    continue;
+                }
+                // OBSERVED, not assumed — the received path already does this. The leaf commits either way,
+                // so a dropped transcript write means the operator's OWN message is missing from their own
+                // transcript with the chain saying it is there, and nothing anywhere said so.
+                if (!this.recordTranscriptMessage(agentName, sessionId, nextExpected, "sent", entry.content, entry.correlationId)) {
+                    this.#logger.error("session.content.released.transcript.failed", {
+                        agentName, sessionId, sequenceNumber: nextExpected, correlationId: entry.correlationId,
+                        impact: "this side's own message is committed to the chain but missing from its transcript",
+                    });
+                }
+            }
+            else if (entry.screenedOut) {
                 this.appendSessionLeaf(agentName, sessionId, "msg", entry.contentHashHex, entry.correlationId);
             }
             else {
-                this.#appendVerifiedContent(agentName, sessionId, entry.content, entry.contentHashHex, senderPubkey, entry.correlationId);
+                this.#appendVerifiedContent(agentName, sessionId, entry.content, entry.contentHashHex, senderPubkey, entry.correlationId, entry.originalContent);
             }
             released++;
             this.#logger.info("session.content.released", {
@@ -4652,10 +5887,35 @@ export class SessionNodeManager {
      */
     async #sendDeliveryAck(agentName, sessionId, contentHash, correlationId) {
         const entry = this.#activeNodes.get(this.#k(agentName, sessionId));
-        if (!entry)
+        if (!entry) {
+            // NOT a silent return. No ACK is exactly this milestone's symptom — the sender's TTF expires
+            // and the message parks — so the one case where we knowingly decline to send one has to say
+            // so, or it is indistinguishable from the defect.
+            this.#logger.debug("content.delivery.ack.skipped", {
+                agentName,
+                sessionId,
+                contentHash: Buffer.from(contentHash).toString("hex"),
+                reason: "session_node_gone",
+                correlationId,
+            });
             return;
+        }
+        // Held outside the try so the catch can retire a stream that was opened and then failed to
+        // write. Without it every failure leaks the OUTBOUND half of the stream the receiver-side
+        // `finally` retires — same defect, other end, other cap. See the note on #handleContentStream.
+        // Assigned IMMEDIATELY after newStream: anything between the two is a window where a throw
+        // leaks the stream because the catch cannot see it.
+        let ackStream;
         try {
             const stream = await entry.node.newStream(entry.counterpartySessionPeerId, CELLO_CONTENT_PROTOCOL_ID);
+            ackStream = stream;
+            // Injected ACK-write failure — thrown from inside the try so it lands in exactly the catch a
+            // real reset lands in, and the whole downstream path (impair → abort → log) runs unmodified.
+            if (this.#ackFaultRemaining > 0) {
+                this.#ackFaultRemaining -= 1;
+                this.#logger.warn("content.delivery.ack.fault.injected", { sessionId });
+                throw new Error("connection_lost: injected delivery-ack fault");
+            }
             const frame = encodeCbor({
                 type: "content_delivery_ack",
                 session_id: sessionId,
@@ -4664,27 +5924,47 @@ export class SessionNodeManager {
                 correlation_id: correlationId,
             });
             stream.send(lp.encode.single(frame));
-            // The receiver-side counterpart to the sender's content.delivery.acked: B has acknowledged
-            // this content `persisted`, so the sender stops retrying/parking. Emitted for BOTH a normally
-            // delivered message AND a terminal-screen block (the block is a definitive receipt — the leaf
-            // is recorded, so the sender must stop) — and deliberately NOT for a transient hold.
+            // NOT SWALLOWED, for the same reason the direct-send path stopped swallowing it: `close()`
+            // waits for the write buffer to drain, so a reset mid-flush throws HERE and that is exactly
+            // the case where the bytes never left. A swallowed close made two things happen at once —
+            // this log claimed the ACK went out while the sender's TTF fired and parked, and the abort in
+            // the catch below (the thing that frees the stream slot) became unreachable.
+            await stream.close();
+            // AFTER the close, because that is when it is true. The receiver-side counterpart to the
+            // sender's content.delivery.acked: B has acknowledged this content `persisted`, so the sender
+            // stops retrying/parking. Emitted for BOTH a normally delivered message AND a terminal-screen
+            // block (the block is a definitive receipt — the leaf is recorded, so the sender must stop) —
+            // and deliberately NOT for a transient hold.
             this.#logger.info("content.delivery.ack.sent", {
                 sessionId,
                 contentHash: Buffer.from(contentHash).toString("hex"),
                 correlationId,
             });
-            try {
-                await stream.close();
-            }
-            catch { /* best-effort close */ }
+            // An agent that mostly LISTENS sends content rarely and ACKs constantly. Clearing only on the
+            // content path would leave exactly those sessions reporting a broken conversation forever
+            // after one bad ACK — the one-way door, on the other send path.
+            this.#clearSessionImpairment(agentName, sessionId, "delivery_ack", correlationId);
         }
         catch (err) {
             this.#logger.warn("content.delivery.ack.send.failed", {
                 sessionId,
                 contentHash: Buffer.from(contentHash).toString("hex"),
                 error: err instanceof Error ? err.message : String(err),
+                // "Cannot write to a stream that is closed" names where the write died, never why. The
+                // why is almost always the per-protocol stream cap, and these two numbers are what turn
+                // that from a log-measurement session into a grep.
+                ...this.#streamCensus(entry.node, entry.counterpartySessionPeerId),
                 correlationId,
             });
+            // The ACK travels the same direct path as our own content, so a failure here is the same
+            // evidence: writes to this counterparty are not landing.
+            this.#markSessionImpaired(agentName, sessionId, { cause: "delivery_ack", error: err instanceof Error ? err.message : String(err), correlationId });
+            if (ackStream !== undefined) {
+                try {
+                    ackStream.abort(err instanceof Error ? err : new Error(String(err)));
+                }
+                catch { /* already gone */ }
+            }
         }
     }
     /** Cancel and drop a single awaiting-ACK entry (e.g. the send failed after arming). */
@@ -4742,9 +6022,17 @@ export class SessionNodeManager {
     // the fragile dependency on that internal timing (review L4).
     async #registerContentHandler(agentName, sessionId, node, _counterpartyPubkey) {
         try {
-            await node.handle(CELLO_CONTENT_PROTOCOL_ID, (stream) => {
-                void this.#handleContentStream(agentName, sessionId, stream);
-            });
+            await node.handle(CELLO_CONTENT_PROTOCOL_ID, (stream, remotePeerId) => {
+                // `.catch` is not decoration: the handler builds its length-prefixed decoder before its own
+                // try, and a throw there would otherwise become an unhandled rejection that takes the
+                // daemon down for one malformed inbound stream.
+                void this.#handleContentStream(agentName, sessionId, stream, remotePeerId).catch((err) => {
+                    this.#logger.warn("session.content.stream.handler.failed", {
+                        sessionId,
+                        error: err instanceof Error ? err.message : String(err),
+                    });
+                });
+            }, { maxInboundStreams: CONTENT_MAX_INBOUND_STREAMS });
         }
         catch (err) {
             this.#logger.error("session.content.handler.register.failed", {
@@ -4930,9 +6218,26 @@ export class SessionNodeManager {
         // No verified position — the caller falls back to the announced hash-dedup path.
         return null;
     }
-    async #handleContentStream(agentName, sessionId, stream) {
-        const iter = lp.decode(stream)[Symbol.asyncIterator]();
+    async #handleContentStream(agentName, sessionId, stream, remotePeerId) {
+        // CLOSING THIS STREAM IS WHAT KEEPS THE SESSION ALIVE PAST ITS 33RD MESSAGE.
+        //
+        // Every content frame and every delivery ACK opens a fresh /cello/content/1.0.0 stream on the
+        // one muxed connection the session holds, and libp2p caps INBOUND streams per protocol per
+        // connection. It enforces that cap AFTER multistream-select has answered, so an over-cap stream
+        // negotiates fine and is reset an instant later, and the SENDER's next `stream.send(...)`
+        // throws "Cannot write to a stream that is closed" — an error that names the exit point and not
+        // one thing about the cause.
+        //
+        // A stream leaves the muxer's set only on its `close` event, and closing our write end triggers
+        // that only once the peer has closed its end too. So a handler that reads its frame and returns
+        // leaves the stream half-open for the life of the connection and the count only ever rises.
+        // Measured on a live daemon: 115 failures over 3.5 hours, with EXACTLY 32 successful streams
+        // before the first one on both affected sessions (M12B Entry 10).
+        //
+        // The decoder is built INSIDE the try so a malformed stream cannot throw past the close below.
+        let iter;
         try {
+            iter = lp.decode(stream)[Symbol.asyncIterator]();
             const result = await iter.next();
             if (result.done || result.value === undefined)
                 return;
@@ -4952,6 +6257,40 @@ export class SessionNodeManager {
                 }
                 return;
             }
+            // DOD-M12B-ABANDON-NOTIFY-1: the counterparty force-abandoned. Handled here, on the same
+            // authenticated stream the delivery acknowledgement rides, and AFTER the session-id check
+            // below cannot be skipped — the frame names its session and the handler is bound to one.
+            if (frame["type"] === "session_abandoned_notice") {
+                // PINNED TO THE COUNTERPARTY. This frame ENDS a conversation, so the stream being
+                // authenticated is not enough — a session node is a promoted standing receiver, and a
+                // standing receiver accepts everyone. libp2p's gater runs at connection establishment and
+                // does not close connections that already exist, so a peer that dialled this node earlier
+                // still holds a live connection after `setAllowedPeer` narrows it. Without this check that
+                // peer could hang up a session it is not party to.
+                //
+                // `remotePeerId` is the Noise-authenticated transport identity, which the handler was
+                // throwing away. Absent means we cannot prove who is speaking, and an unprovable claim to
+                // end a session is refused.
+                const expected = this.#activeNodes.get(this.#k(agentName, sessionId))?.counterpartySessionPeerId;
+                if (!remotePeerId || !expected || remotePeerId !== expected) {
+                    this.#logger.warn("session.content.peer_mismatch", {
+                        sessionId, frameType: "session_abandoned_notice",
+                        remotePeerId: remotePeerId ?? "(absent)", expected: expected ?? "(unknown)",
+                    });
+                    return;
+                }
+                // REQUIRED and equal — absence is not a pass. The frame names its session and the handler
+                // is bound to one; treating a missing field as agreement is how a guard stops guarding.
+                const claimed = frame["session_id"];
+                if (typeof claimed !== "string" || claimed !== sessionId) {
+                    this.#logger.warn("session.content.session_mismatch", {
+                        sessionId, claimedSessionId: typeof claimed === "string" ? claimed : "(absent)",
+                    });
+                    return;
+                }
+                void this.retireOnCounterpartyAbandon(agentName, sessionId, correlationId);
+                return;
+            }
             if (frame["type"] !== "content_frame") {
                 // LOGGED, not silently dropped. This handler is bound to one session, and a frame it does
                 // not understand arriving on that stream is either a peer speaking a newer protocol or a
@@ -5026,6 +6365,48 @@ export class SessionNodeManager {
                 error: err instanceof Error ? err.message : String(err),
             });
         }
+        finally {
+            // `close()` waits only for OUR write buffer, which is empty here, so this cannot stall the
+            // handler; it runs on every exit above, and there are several early returns.
+            try {
+                await stream.close();
+            }
+            catch (err) {
+                // NOT SILENT. A close that fails here is the signature of the cap biting from the other
+                // side, and it was the absence of exactly this line that turned the original diagnosis
+                // into a 6,451-record log measurement.
+                this.#logger.warn("session.content.stream.close.failed", {
+                    sessionId,
+                    error: err instanceof Error ? err.message : String(err),
+                });
+                try {
+                    stream.abort(err instanceof Error ? err : new Error(String(err)));
+                }
+                catch { /* already gone */ }
+                return;
+            }
+            // OUR CLOSE ALONE DOES NOT FREE THE SLOT — the peer has to close its end too, and a peer
+            // owns its own daemon. Without this, someone who opens content streams and never closes them
+            // pins every inbound slot we have and puts us straight back into the defect above, with the
+            // same unreadable error. `abort` resets unilaterally, so it works regardless of the peer;
+            // the delay is what keeps it from landing while a well-behaved sender is still inside its
+            // own `close()`. Unref'd so it can never hold the process open at shutdown, and tracked so
+            // teardown can drop it.
+            if (stream.status === "open" || stream.status === "closing") {
+                const linger = setTimeout(() => {
+                    this.#lingeringStreams.delete(linger);
+                    if (stream.status !== "open" && stream.status !== "closing")
+                        return;
+                    this.#logger.debug("session.content.stream.linger.reset", { sessionId });
+                    try {
+                        stream.abort(new Error("inbound content stream not closed by peer"));
+                    }
+                    catch { /* already gone */ }
+                }, CONTENT_STREAM_LINGER_MS);
+                linger.unref?.();
+                this.#lingeringStreams.add(linger);
+            }
+        }
     }
     /**
      * M7-SESSION-001 AC-004/AC-005: Register a relay stream for an active session.
@@ -5681,13 +7062,85 @@ export class SessionNodeManager {
         return flipped;
     }
     /** @returns true iff the UPDATE was executed without error (a failed write is logged, never thrown). */
-    #updateSessionStatus(agentName, sessionId, status) {
+    /**
+     * DOD-M12B-STRAND-1 — move a terminal session's held frames to the annex.
+     *
+     * A held frame is content this agent RECEIVED and VERIFIED. When its session ends it can never
+     * join that chain (appending behind a committed root is not an option, and ingest refuses a
+     * terminal session outright), but it is still the operator's mail and no other copy exists —
+     * the sender was never acknowledged for it. `sealed_session_annex` is where M12-P17 already puts
+     * content that arrives for an ended session; this is the same content arriving slightly earlier.
+     *
+     * ANNEX FIRST, DELETE SECOND, per row. A crash between them costs a duplicate the annex's
+     * INSERT OR IGNORE absorbs; the other order costs the message. A row whose annex write fails is
+     * KEPT — the retention sweep will find it again, and a leftover row is cheaper than a lost one.
+     */
+    #annexHeldContentOnTerminal(agentName, sessionId, status) {
+        if (!this.#db)
+            return;
+        let rows;
+        try {
+            rows = this.#db.prepare(`SELECT canonical_seq, content_blob, content_hash_hex, held_at, origin
+           FROM held_content WHERE agent_id = ? AND session_id = ? ORDER BY canonical_seq ASC`).all(this.#requireAgentId(agentName), sessionId);
+        }
+        catch (err) {
+            this.#logger.error("session.content.held.annex.scan.failed", {
+                agentName, sessionId, status,
+                impact: "held frames for a terminal session were not moved to the annex and remain unreadable",
+                error: err instanceof Error ? err.message : String(err),
+            });
+            return;
+        }
+        if (rows.length === 0)
+            return;
+        const counterparty = this.getSessionRecord(agentName, sessionId)?.counterparty_pubkey ?? null;
+        let annexed = 0;
+        let kept = 0;
+        for (const row of rows) {
+            // DOD-M12B-INDEX-1: ATTRIBUTION. A `sent` row is our own message. Stamping the counterparty's
+            // pubkey on it would put our words in their mouth in the one record that survives the
+            // session — the same failure the release path was changed to avoid, on the drain it did not
+            // touch. `#ownPubkeyHex` is null only when the identity cannot be resolved, and a null sender
+            // reads as "unattributed", which is true, rather than as a false attribution.
+            const sender = row.origin === "sent" ? this.#ownPubkeyHex(agentName) : counterparty;
+            if (this.recordSealedAnnex(agentName, sessionId, row.content_hash_hex, new Uint8Array(row.content_blob), sender)) {
+                this.#deleteHeldContent(agentName, sessionId, row.canonical_seq);
+                // AND OUT OF THE IN-MEMORY MAP. Measured live 2026-08-17 on daemon 0.0.170: this frame is
+                // now safe in the annex and its durable row is gone — but teardown still found it in the
+                // map, counted `held_content` for the session, got 0, and fired
+                // `session.content.held.lost`: "verified content was destroyed". Ten frames were annexed
+                // and the same ten were reported destroyed, in the same second. A false alarm on the most
+                // serious event in the system is worse than no alarm, because the next investigation goes
+                // looking for content that was never lost.
+                this.#heldContent.get(this.#k(agentName, sessionId))?.delete(row.canonical_seq);
+                annexed++;
+            }
+            else {
+                kept++;
+            }
+        }
+        this.#logger.warn("session.content.held.annexed", {
+            agentName, sessionId, status, annexed, kept,
+            // The consumer of `held_at`: how long the oldest frame waited before its session ended.
+            oldestHeldMs: Date.now() - Math.min(...rows.map((r) => r.held_at)),
+            impact: "these messages arrived and verified but never joined the chain — they are readable from the annex, not the transcript",
+        });
+    }
+    #updateSessionStatus(agentName, sessionId, status, 
+    // DOD-CAP-SELF-HEAL-1: who caused an interruption, when this call is the one causing it.
+    // Omitting it leaves the column NULL, which the acceptance bound reads as the counterparty's —
+    // so a LOCAL teardown that forgets to say so is charged to the peer. That is exactly how the
+    // operator's own kill switch (`cello_set_agent_offline` → destroySessionNode) was locking out
+    // a counterparty who had done nothing.
+    interruptedBy) {
         if (!this.#db)
             return false;
         const now = Date.now();
         try {
             const res = this.#db
-                .prepare("UPDATE sessions SET status = ?, updated_at = ? WHERE agent_id = ? AND session_id = ?")
+                .prepare(interruptedBy === undefined
+                ? "UPDATE sessions SET status = ?, updated_at = ? WHERE agent_id = ? AND session_id = ?"
+                : "UPDATE sessions SET status = ?, updated_at = ?, interrupted_by = 'local' WHERE agent_id = ? AND session_id = ?")
                 .run(status, now, this.#requireAgentId(agentName), sessionId);
             // "Did not throw" is NOT "landed". An UPDATE whose WHERE matches no row — a wrong agent_id, a
             // session_id with no row — succeeds silently and changes nothing. Reporting that as a written
@@ -5708,6 +7161,14 @@ export class SessionNodeManager {
             // is still, on disk, drainable. 'interrupted' and 'seal_interrupted_pending' are deliberately
             // NOT terminal — both can still complete, and reaping them would destroy live content.
             if (status === "sealed" || status === "abandoned") {
+                // DOD-M12B-STRAND-1: held frames outlive the chain that could have carried them.
+                //
+                // Once a session is terminal, `ingestReceivedContent` refuses it — and #releaseHeld is only
+                // reachable from ingest — so no code path that exists can ever release a held frame again.
+                // Left alone the rows sit on disk, unreachable by any surface, while the teardown alarm
+                // reports `lost: 0`: a success message for content that has just become permanently
+                // unreadable. The annex is the store built for exactly this shape.
+                this.#annexHeldContentOnTerminal(agentName, sessionId, status);
                 try {
                     this.#onSessionTerminal?.(sessionId, status);
                 }