diff --git a/.claude/board/ISSUES.md b/.claude/board/ISSUES.md
index c10737cb7..d7eeb091b 100644
--- a/.claude/board/ISSUES.md
+++ b/.claude/board/ISSUES.md
@@ -1,6 +1,6 @@
 ## ISS-SUPERVISOR-RESPAWN-BACKOFF-NOT-APPLIED — a crashed consumer actor is respawned at once and the respawn is rescheduled forever (2026-10-13)
 
-**Status:** OPEN. **Basis:** VERIFIED-IN-CODE (`crates/lance-graph-supervisor/src/supervisor.rs`, `RespawnG` arm); found by the codex review of #1536.
+**Status:** RESOLVED 2026-10-13 by #1540 (merged `f2bf199`). The delayed cast is now the only respawn path and is scheduled from `ActorTerminated`; the `RespawnG` arm spawns once and does nothing if the slot is already live; the crash count is carried and restarts at 1 when the previous crash is more than five minutes old (spec §5.3); more than 10 crashes in that window escalate and the child stays down. TEST-PINNED in `tests/supervisor_respawn_backoff.rs` (one respawn per crash after its backoff; escalation; window reset), each disable-verified red. The bullets below describe the defect as found. **Was:** OPEN. **Basis:** VERIFIED-IN-CODE (`crates/lance-graph-supervisor/src/supervisor.rs`, `RespawnG` arm); found by the codex review of #1536.
 
 - On `ActorTerminated` the supervisor casts `RespawnG { crash_count: n }`. The arm computes the backoff delay, spawns a task that sleeps that delay and casts `RespawnG { crash_count: 0 }`, and then calls `spawn_consumer_actor` right away (the comment says the delayed cast is a retry "if we need to").
 - The delayed `RespawnG` runs the same arm: another immediate spawn, another delayed cast. So the backoff never delays the first respawn, the crash count is reset to 0 on every round, and each round schedules the next one 100 ms later, whether or not the child is alive.
diff --git a/.claude/board/LATEST_STATE.md b/.claude/board/LATEST_STATE.md
index 663eac6f5..0642053e0 100644
--- a/.claude/board/LATEST_STATE.md
+++ b/.claude/board/LATEST_STATE.md
@@ -1,4 +1,4 @@
-## 2026-10-13 — main after #1505, #1506, #1507–#1509, #1514, #1516, #1518, #1520, #1523, #1524, #1525, #1526, #1529, #1530, #1531, #1534, #1536, #1537 (merged)
+## 2026-10-13 — main after #1505, #1506, #1507–#1509, #1514, #1516, #1518, #1520, #1523, #1524, #1525, #1526, #1529, #1530, #1531, #1534, #1536, #1537, #1538, #1539, #1540 (merged)
 
 - **On `main`, live:** `.cargo/config.toml` defaults to `-Ctarget-cpu=native` with `config-v3.toml` as the portable baseline and `debug = 0` everywhere (#1505); `compute_dag_kahn` checks all 256 field positions (#1506); R2IL `MIN`/`MAX` lowering onto mask-risc (#1518, the part outside glove-parity).
 - **Merged then withdrawn by #1524 (D-CMR-RESET):** the class-method layer of `lance-graph-glove-parity` (`method`, `call`, `decl`, `gremlin`, `awareness`; #1507, #1508, #1509, #1514, #1516, #1518, #1520). Prior state: `backup/session-011BMJ-2026-10-12`. Inventory: entry `2026-10-12-session-reset-inventory.md`.
@@ -8,7 +8,9 @@
 - **On `main` (#1530 `c864721`, #1531 `86e3174`):** the thinking-engine, cognitive-stack and symbiont images build on `rust:1.99.0-bookworm` against ndarray `master`; the thinking-engine images build v4 with `--target <host triple>`.
 - **On `main` (#1534, merged `0a6708b`):** `lance-graph-r2il` is a workspace member; r2il, `lance-graph-nal` and `r2il-mask-abi-probe` take OGAR by git `main`, so the graph holds one `ogar-loco` and one `lance-graph-contract` (D-FOLD-CORE-a).
 - **On `main` (#1537, merged `befbf1d`):** core `lance-graph` depends on mask-risc, Quack and R2IL by default and exposes them as `lance_graph::fold` (D-FOLD-CORE-b). NAL stays in `crates/lance-graph-nal` (D-FOLD-CORE-c dropped).
-- **Docs only (#1536, merged `2b28a19`):** outer-membrane graph lab, epoch-indexed temporal graph agenda, and the ONA → NAL/R2IL fold-method program (W3–W9). Proposals; nothing in them is built by that PR. The supervisor's restart backoff is not applied (`ISS-SUPERVISOR-RESPAWN-BACKOFF-NOT-APPLIED`).
+- **Docs only (#1536, merged `2b28a19`):** outer-membrane graph lab, epoch-indexed temporal graph agenda, and the ONA → NAL/R2IL fold-method program (W3–W9). Proposals; nothing in them is built by that PR.
+- **On `main` (#1538, merged `4ceaab3`):** D-NALP-3, `nal::Revision`: evidence revision as one grouped fold over stored `w`/`f`, refusing overlapping stamps and evidence without provenance. Where ONA saturates mid-fold and its answer depends on arrival order, the fold keeps the exact sum (pinned against the ONA oracle).
+- **On `main` (#1540, merged `f2bf199`):** the supervisor applies its restart backoff; `ISS-SUPERVISOR-RESPAWN-BACKOFF-NOT-APPLIED` resolved. #1539 (`3635109`) corrected #1536's docs and board.
 - **CI (#1526, merged `abc0a1f`):** the native default plus `rust-cache` reused AVX-512 artifacts on runners without them (SIGILL in `dir-sim log in Lance`). `native` is now written out as each runner's exact feature list, and the cache is keyed to it.
 - **Rules carried from this session (operator):** classid-dependent methods go through the existing `hotplug.rs` plug-and-play; NAL/NARS operations are global core methods (no classid check, no plug, no `BeliefClass`); internal pins are forbidden (the v3 pins in the Dockerfiles and wheel build were replaced in #1529).
 
diff --git a/.claude/board/PR_ARC_INVENTORY.md b/.claude/board/PR_ARC_INVENTORY.md
index e39f41785..f5cf87efe 100644
--- a/.claude/board/PR_ARC_INVENTORY.md
+++ b/.claude/board/PR_ARC_INVENTORY.md
@@ -1,3 +1,17 @@
+## 2026-10-13 — lance-graph PR #1540 (merged `f2bf199`, branch `ccr-9f3b8075-74u47l-supervisor-backoff`, head `dd3ab5d`, 5 commits) — supervisor: apply the restart backoff
+
+- **Added:** `ActorTerminated` computes the crash count (carried from the slot; restarts at 1 when the previous crash is more than `CRASH_WINDOW` = 5 min old), escalates past 10 crashes, otherwise sleeps `backoff_delay(n) = min(100 ms · 2^n, 30 s)` and casts the only `RespawnG`. The arm spawns once, skips a slot that is already live, and stores the count and crash time on the new slot. `tests/supervisor_respawn_backoff.rs` (paused tokio time): one respawn per crash after its backoff, crash 11 stays down, the count restarts after a quiet window; each guard disable-verified red.
+- **Locked:** before, the arm spawned immediately and also scheduled a `RespawnG { crash_count: 0 }` that re-ran itself every 100 ms forever, so neither backoff nor escalation could take effect. The crash time uses `tokio::time::Instant` (a `std` instant ignores paused time, which hid the window bug from the test).
+- **Deferred / open:** none; `ISS-SUPERVISOR-RESPAWN-BACKOFF-NOT-APPLIED` resolved.
+- **Confidence:** high; three tests, disable runs red, CI green.
+
+## 2026-10-13 — lance-graph PR #1538 (merged `4ceaab3`, branch `ccr-9f3b8075-74u47l-nal-w3`, head `12f3291`, 5 commits) — D-NALP-3: NAL revision as one grouped fold
+
+- **Added:** `nal::Revision` in `crates/lance-graph-nal`: groups evidence by the class its stamps define and folds stored `w`/`f` with `GroupCrossPowerSumsI32`; refuses overlapping stamps and evidence without provenance. `tests/w3.rs`, 11 tests against a scalar oracle and ONA pinned at `dc4efd0`, and the revision points OpenNARS printed for `revision_points.nal`.
+- **Locked:** the fold computes the exact `w = Σwᵢ` and caps only the result. ONA caps pairwise, so once evidence saturates its answer depends on arrival order (measured on one fixture: 104 positive-first, 151 negative-first; the fold gives 128 either way). Documented as a divergence and pinned.
+- **Deferred / open:** W4–W9 of the ONA program (per-row truth laws, inference over populations, the `NAL::narsese()` surface).
+- **Confidence:** high; disable runs red, oracle parity up to ONA's mid-fold saturation.
+
 ## 2026-10-13 — lance-graph PR #1536 (merged `2b28a19`, branch `docs/outer-membrane-graph-lab-20261013`, head `93b0993`, 5 commits) — docs: outer membrane graph lab, epoch-indexed temporal graph agenda, ONA → NAL/R2IL fold-method program
 
 - **Added:** three proposal documents under `docs/architecture/`: `OUTER_MEMBRANE_GRAPH_LAB_VERTICAL_SLICE.md` (ownership rule, callcenter/supervisor evidence table, a falsifiable two-node one-edge vertical slice), `EPOCH_TEMPORAL_GRAPH_RESEARCH_AGENDA.md` (E01–E20 exploration questions, three time axes, T0–T6 sequence), `ONA_NAL_R2IL_FOLD_METHOD_ASSIMILATION.md` (ONA source map, W3–W9 coding waves, falsifiers). No code.
diff --git a/.claude/board/STATUS_BOARD.md b/.claude/board/STATUS_BOARD.md
index 08c15b258..64841a178 100644
--- a/.claude/board/STATUS_BOARD.md
+++ b/.claude/board/STATUS_BOARD.md
@@ -92,7 +92,7 @@ Plan: `.claude/plans/nal-polyfill-v1.md`. Template: D-SQLP. No opcode, tenant or
 | **D-NALP-0** | Plan, board rows | In PR | docs only |
 | **D-NALP-1** | Statement layout + declared ClassView reading; additive `WideFieldMask` projection path on `ClassView`; binder's mask; term width decision | Partial: #1525 merged `08635ed` (NodeGuid reading `nal-statement-v1`, term width u16, aperture selection); ClassView wide projection and binder mask Queued | only the mask's lanes are read; undeclared reading and masked-out field refuse; Small = Wide answers |
 | **D-NALP-2** | Query procedures: select, truth threshold, choice by confidence, support count | Shipped #1525 `08635ed` (`crates/lance-graph-nal`: NodeGuid key, byte apertures; 7 oracle tests, 7 disable runs red; 19–248 ns at 1,048,576 rows); support count and argmax not built | equal to a row-at-a-time oracle; one-pattern threshold disable red; no new opcode |
-| **D-NALP-3** | Revision as `GroupCrossPowerSumsI32` over stored `w`/`f` | In PR (`nal::Revision`, key = evidence class from the stamps; refuses overlap and missing provenance; 10 tests, 7 disable runs red) | equal to exact integer revision; OpenNARS revision points; duplicate evidence counts once |
+| **D-NALP-3** | Revision as `GroupCrossPowerSumsI32` over stored `w`/`f` | Shipped #1538 `4ceaab3` (`nal::Revision`, key = evidence class from the stamps; refuses overlap and missing provenance; 11 tests, 7 disable runs red; ONA's mid-fold saturation is a documented divergence, the fold keeps the exact sum) | equal to exact integer revision; OpenNARS revision points; duplicate evidence counts once |
 | **D-NALP-4** | Per-row truth law into a lane: `MapPairU8` vs fixed-point multiply-high | Queued | insufficiency proof; per-row parity with `isa::truth`, error in codes; one shape also serves SQL nullable arithmetic |
 | **D-NALP-5** | Deduction / induction / abduction / analogy over populations | Queued (after D-NALP-4) | equals pairwise scalar rule; abduction direction test red for the reversed statement |
 | **D-NALP-6** | `lance-graph-nal`: typed `nal::{Operation}` surface (`NalOp`), parser (tokens only), binder over the reverse label index, `NAL::narsese()`, typed refusals | Queued | text and typed calls give byte-identical `Program`s; no string in the plan, rows or wire; ambiguous / unknown label refuses; lowers from the bound plan; no fallback; no `Program` outside procedures |
diff --git a/docs/architecture/OUTER_MEMBRANE_GRAPH_LAB_VERTICAL_SLICE.md b/docs/architecture/OUTER_MEMBRANE_GRAPH_LAB_VERTICAL_SLICE.md
index 0c5819ad4..a5e3229a8 100644
--- a/docs/architecture/OUTER_MEMBRANE_GRAPH_LAB_VERTICAL_SLICE.md
+++ b/docs/architecture/OUTER_MEMBRANE_GRAPH_LAB_VERTICAL_SLICE.md
@@ -60,7 +60,7 @@ The existing [`ExternalMembrane`](../../crates/lance-graph-callcenter/src/lance_
 | Capability | Current evidence | Confidence / gap |
 |---|---|---|
 | Per-consumer actor slots, health | `supervisor.rs` | Implemented skeleton; confirm runtime integration before claiming production behavior |
-| Restart backoff | `RespawnG` arm in `supervisor.rs` | **Not applied.** The arm schedules a delayed `RespawnG` with `crash_count: 0` and then calls `spawn_consumer_actor` immediately; the delayed message repeats both, so respawns start without waiting and keep being rescheduled every 100 ms (`ISS-SUPERVISOR-RESPAWN-BACKOFF-NOT-APPLIED`) |
+| Restart backoff | `RespawnG` arm in `supervisor.rs` | Applied (#1540). On `ActorTerminated` the supervisor waits `min(100 ms · 2^n, 30 s)` and then sends the only `RespawnG`; the count is carried to the new child, restarts at 1 after five quiet minutes, and more than 10 crashes in five minutes leave the child down (`ISS-SUPERVISOR-RESPAWN-BACKOFF-NOT-APPLIED`, resolved) |
 | Default mailbox target `1024` / per-module override | `DEFAULT_MAILBOX_CAPACITY`, `ModuleEntry.mailbox_capacity` | **Declared**, but `spawn_consumer_actor` currently calls `Actor::spawn_linked` without applying that capacity |
 | Normal `DispatchToG` forwarding | `supervisor.rs::handle` | **Not implemented** for non-`Health` envelopes: returns `DispatchNotImplemented` |
 | Inbound `DrainTask` queue | `callcenter/src/drain.rs` | **Unbounded** Tokio mpsc in realtime build, even though child mailbox backpressure is a design goal |