feat(fleet): add relay custody and resilient transport

Author: issdandavis

docs/SYSTEMS_CATALOG.md

@@ -62,6 +62,10 @@ _Verified 2026-06-16: `scbe -V` → `scbe 4.2.1`; CLI smoke green; wheel install
 - **M-TEF research packet** — `docs/research/mtef_research_compendium_2026-06-17.md`
   and the linked PDF preserve the Magneto-Triboelectric Fluid Cell literature synthesis.
   Treat it as concept/prototype evidence, not proof of a working device.
+- **Magnetic pocket routing / wafer boundary** —
+  `docs/specs/MAGNETIC_POCKET_ROUTING_WAFER_RESEARCH_20260919.md` maps the existing
+  magnetic-logic, ferrofluid-control, and wafer-router work to published physical-computing
+  results and a gated tabletop-to-chip validation ladder.
 - **Space life-support animals packet** —
   `docs/research/space_life_support_animals_husbandry_energy_2026-06-17.md` captures the
   animal-husbandry / micro-energy idea as an Idea Seed with required source, welfare,

docs/specs/FLEET_COMPOSITION_COORDINATION.md

@@ -87,10 +87,13 @@ Every governance check and every candidate assessment is billed. This avoids
 the execution-accounting error where a router evaluates many candidates but
 charges one operation.
 
-The stress matrix uses nominal links, delay and reorder, duplicate delivery,
-loss with custody retransmission, blackout with custody, and permanent loss
-without custody. Duplicate bundles are deduplicated. Custody-backed loss must
-converge; permanent loss without custody must report divergence.
+The stress matrix uses seven conditions: nominal links, delay and reorder,
+duplicate delivery, loss with custody retransmission, a round-trip-dependent
+blackout control, the same blackout with pre-synchronized local autonomy and
+store-carry-forward custody, and permanent loss without custody. Duplicate
+bundles are deduplicated. Custody-backed loss must converge; permanent loss
+without custody must report divergence. Every eighth task is marked as truly
+remote and still requires the live link in both blackout conditions.
 
 ## Measured result
 
@@ -100,44 +103,107 @@ Command:
 npm run benchmark:fleet-composition -- --tasks 180 --seeds 7,11,19,23,31,43,59,71,83
 ```
 
-This produces 270 runs: nine seeds, six network conditions, and five policies.
+This produces 315 runs: nine seeds, seven network conditions, and five policies.
 
 | Policy | Completion | Cost per completed task | Mean transition stability | Churn | Capability misses |
 | --- | ---: | ---: | ---: | ---: | ---: |
-| sticky | 0.7943 | 77.9297 | 0.3756 | 0.0168 | 1.8704 |
-| distance | 0.7923 | 78.6584 | 0.0500 | 0.0056 | 2.8333 |
-| composition-aware | 0.7305 | 46.7603 | 0.6860 | 0.5134 | 0.0000 |
-| stability-guarded | 0.7869 | 61.1286 | 0.8360 | 0.2484 | 0.0000 |
-| round-robin | 0.4195 | 133.7831 | 0.5831 | 0.8994 | 67.1296 |
+| sticky | 0.7543 | 126.6063 | 0.3756 | 0.0168 | 1.8889 |
+| distance | 0.7525 | 128.1181 | 0.0500 | 0.0056 | 2.8571 |
+| composition-aware | 0.7014 | 79.1857 | 0.6871 | 0.5063 | 0.0000 |
+| stability-guarded | 0.7508 | 99.8604 | 0.8302 | 0.2489 | 0.0000 |
+| round-robin | 0.4113 | 277.5470 | 0.5831 | 0.8994 | 67.8254 |
 
 The aggregate claim is **NO_LIFT** under the repository rule requiring the
 custom arm's completion delta to exceed twice pooled sample standard deviation
 against both controls. Stability guarding reduced cost per completed task by
-about 21.6% versus sticky and 22.3% versus distance, while completion was 0.74
-and 0.53 percentage points lower, respectively. Those cost effects are
-underpowered across the mixed scenarios. Compared with unguarded adaptive
-routing, the guard cut churn by about 51.6% and raised mean transition stability
-from 0.6860 to 0.8360.
+about 21.1% versus sticky and 22.1% versus distance, while completion was 0.35
+and 0.17 percentage points lower, respectively. Those aggregate cost effects
+remain underpowered across the mixed scenarios. Compared with unguarded
+adaptive routing, the guard cut churn by about 50.8% and raised mean transition
+stability from 0.6871 to 0.8302.
 
 The measured result supports continued engineering, not a superiority claim.
 
-Across nominal, delay/reorder, duplicate, and loss-with-custody scenarios, the
-guarded policy completed 0.9358 of tasks at a cost near 50.65 per completion;
-the two controls completed roughly 0.9370 to 0.9395 at costs from 66.22 to
-66.84. Blackout exposed the largest completion gap: 0.4247 for the guarded arm
-versus 0.4549 to 0.4568 for the controls. Under permanent loss without custody,
-all policies correctly reported state divergence. Scenario-specific effects and
-their uncertainty gates are retained in the JSON benchmark receipt.
+Under permanent loss without custody, all policies correctly reported state
+divergence. Scenario-specific effects and their uncertainty gates are retained
+in the JSON benchmark receipt.
+
+## Mars relay blackout ablation
+
+The blackout comparison changes one architectural assumption while keeping the
+task stream, loss, delay, reordering, custody, seeds, and policies paired. In the
+control, every task waits for the remote round trip. In the relay condition,
+157 of 180 tasks per seed are already authorized for local execution and their
+receipts are buffered; the remaining 23 tasks still need the remote link.
+
+| Policy | Round-trip completion | Relay completion | Completion delta | Round-trip cost | Relay cost |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| sticky | 0.1321 | 0.8389 | +0.7068 | 481.0277 | 74.2911 |
+| distance | 0.1315 | 0.8370 | +0.7056 | 488.0270 | 74.9010 |
+| composition-aware | 0.1148 | 0.7815 | +0.6667 | 326.9579 | 42.4208 |
+| stability-guarded | 0.1253 | 0.8333 | +0.7080 | 387.9293 | 57.0642 |
+| round-robin | 0.0395 | 0.4759 | +0.4364 | 1362.4905 | 101.9915 |
+
+Every policy cleared the two-pooled-standard-deviation threshold for completion,
+cost, and deadline misses, so the architecture ablation is **SUPPORTED** in this
+simulation. For the stability-guarded policy, mean deadline misses fell from
+156.0 to 19.56. This is evidence for pre-synchronized local autonomy plus
+custody, not for a physical Mars link. The 14 blackout ticks are normalized
+stress units; they do not claim fourteen Earth days, RF throughput, hardware
+reliability, or deployment security.
+
+## Multiplexed transport boundary
+
+The six tongues, harmonic frequencies, and reversible views can form a useful
+transport braid when their jobs stay explicit:
+
+1. Assemble commands into the canonical semantic opcode tape.
+2. Compress repeated structure with a declared lossless codec or shared
+   codebook.
+3. Authenticate and encrypt with a standard signed/AEAD or PQC envelope.
+4. Add erasure coding, then assign shards to tongue/frequency lanes.
+5. Treat forward, reverse, complement, palindrome, spatial, or audio forms as
+   bijective representations and independent checks only when exact decoding is
+   tested.
+6. Reconstruct, authenticate, govern, and only then execute.
+
+The repository already proves byte-to-token round trips in all six tongues,
+compact semantic-opcode assembly, and a separate FSK audio prototype. These
+pieces do not create free channel capacity. Ten arbitrary commands still require
+enough source entropy unless both ends share a codebook. A palindrome repeats
+the same symbols; two independent meanings require orientation-specific
+decoders or additional symbols. Spatial negative-space readings are suitable
+for human display or steganographic experiments, but not as an authoritative
+machine-command lane because ordinary reformatting can destroy them.
+
+The first deterministic fixture is implemented in
+src/fleet/bijective_frequency_transport.py and
+scripts/benchmark/bijective_frequency_transport.py. Ten semicolon-separated
+commands occupied 97 UTF-8 bytes, assembled to 47 semantic-opcode bytes, and
+compressed to 44 bytes. Five data shards plus one parity shard used 54 wire
+symbols, or 1.2273 times the packed payload. That was 79.55% fewer symbols than
+sending six complete packed copies. All six possible single-lane erasures and
+all six single-lane corruptions reconstructed the exact opcode tape; all fifteen
+two-lane erasures were rejected; all six tongue views round-tripped exactly.
+
+Those are structural software results for one deterministic fixture. The next
+benchmark must add multiple payload distributions, bit-error and burst-loss
+channel traces, decode latency, outer-envelope overhead, and standard erasure
+codes. Security claims remain attached to the standard authenticated envelope,
+not to geometric or linguistic transforms. Physical RF/audio performance is
+still untested.
 
 ## Production path
 
-1. Stratify claims by network condition instead of relying only on a mixed
-   aggregate.
-2. Add short-horizon planning so the router can price several upcoming tasks
+1. Bind the Mars relay abstraction to contact-window queues, finite buffers,
+   relay-node loss, and delayed receipt conflicts.
+2. Benchmark the six-lane command braid against raw, repetition, and standard
+   erasure-coded controls over the same noisy channel traces.
+3. Add short-horizon planning so the router can price several upcoming tasks
    rather than react one task at a time.
-3. Calibrate normalized costs with live model token, latency, queue, and dollar
+4. Calibrate normalized costs with live model token, latency, queue, and dollar
    measurements.
-4. Bind decisions to the signed fleet decision envelope and existing TypeScript
+5. Bind decisions to the signed fleet decision envelope and existing TypeScript
    governance gate.
-5. Validate against real Redis/DTN transport, concurrent workers, and process
+6. Validate against real Redis/DTN transport, concurrent workers, and process
    failure before using the router as a control-plane component.

docs/specs/MAGNETIC_POCKET_ROUTING_WAFER_RESEARCH_20260919.md

@@ -0,0 +1,152 @@
+# Magnetic Pocket Routing and Wafer Prototype Boundary
+
+**Status:** research specification; software evidence exists, physical device unbuilt
+
+**Date:** 2026-09-19
+
+**Scope:** magnetic pocket nodes, field-directed routing, fluid actuation, and the Clay wafer design
+
+## Engineering interpretation
+
+The useful form of the concept is a closed-loop physical state machine. A goal does not
+directly move matter. Governance converts an explicit goal and measured state into a bounded
+control vector; coils or electrodes create a field; physical nodes move or switch; sensors
+measure the result; and the controller accepts, corrects, or rejects the transition.
+
+| Design image | Engineering variable |
+|---|---|
+| pocket node | droplet trap, domain-wall site, nanomagnet, or skyrmion confinement region |
+| subtle magnetic shift | bounded coil current, voltage pulse, or bias-field change |
+| system intention | explicit goal vector plus policy constraints |
+| lever or ball bearing | mobile droplet, domain wall, or other measured state carrier |
+| dense/loose fluid | field-dependent flow or yield response; material must be specified |
+| wafer neighborhood | graph edge implemented by a wire, field coupling, waveguide, or software route |
+
+A minimal controller can be written as
+
+\[
+\mathbf B(\mathbf x,t)=\sum_i u_i(t)\,\mathbf b_i(\mathbf x),
+\qquad
+\mathbf u_t^*=\arg\min_{\mathbf u\in\mathcal U}
+\left\|F(\mathbf s_t,\mathbf u)-\mathbf s_{goal}\right\|_Q^2
++\lambda\|\mathbf u\|_2^2,
+\]
+
+where `s_t` is the sensed physical state, `u` is a bounded actuator command, `F` is a
+measured transition model, and `U` contains current, temperature, timing, and policy limits.
+This formalizes "nodes drifting with intention" without giving an unmeasured internal state
+direct control of hardware.
+
+## What is already supported locally
+
+1. `C:\dev\clay-wafer-net\clay_wafer_net\magnetic_logic.py` demonstrates the logical
+   boundary of a threshold-coupled magnetic model. Three ferromagnetic inputs implement a
+   monotone majority function. Antiferromagnetic coupling plus bias implements NAND and NOR.
+   A single threshold node cannot implement XOR; a two-layer construction does so exactly.
+   Clocking and fabrication remain outside that proof.
+2. `C:\dev\ferro_field.py` maps a small command grammar into field surfaces and rejects an
+   unknown operation. It is a control-language simulation, not a material or device result.
+3. `C:\dev\ferrofluid_habitat_control_SPEC.md` already separates controllable magnet arrays
+   and simulated surfaces from speculative integrated hardware.
+4. `C:\dev\tongue-compiler\artifacts\wafer_nvidia_review_20260912\REVIEW.md` found a real
+   64-expert/top-4 software router and later repair evidence found finite router/gate gradients.
+   It found no CUDA kernel, RTL, physical layout, or device benchmark. The wafer is presently a
+   neural routing architecture.
+
+## What published experiments establish
+
+- Rotating magnetic fields and patterned tracks have moved ferrofluid droplets through AND,
+  OR, XOR, NOT, and NAND gates, fanout, a full adder, a flip-flop, and a finite-state machine.
+  This strongly supports a tabletop routing demonstrator, but the reported platform operated
+  at fluidic rather than processor-clock timescales.
+- A water-based ferrofluid has experimentally shown memristive behavior, short- and long-term
+  memory, and reservoir-computing classification with electrical programming and RF readout.
+- Nanomagnetic logic chains have shown sub-nanosecond signal propagation under explicit
+  clocking. Domain-wall prototypes have shown inversion, buffering, gain, fanout, and cascaded
+  room-temperature operation.
+- Geometrically confined magnetic skyrmions have performed physical reservoir computing. Their
+  nonlinear trajectories and relaxation toward a stable region closely match the "pocket
+  node" intuition, although the demonstrated systems require specialized thin films and
+  readout hardware.
+
+These are adjacent results. None demonstrates the full SCBE/Clay wafer.
+
+## Material choice matters
+
+Do not treat ferrofluid, magnetorheological fluid, and a spintronic thin film as interchangeable.
+
+| Medium | Best fit here | Main limitation |
+|---|---|---|
+| ferrofluid droplets | visible routing, logic, memory experiments, education | slow, surface/flow variation, bulky sensing |
+| magnetorheological fluid or elastomer | variable damping, stiffness, clutching, and actuation | thermal and mechanical response is far slower than digital logic |
+| nanomagnets/domain walls | nonvolatile Boolean logic and memory | clocking, fabrication variation, write current, fanout |
+| skyrmions/spin waves | nonlinear reservoirs and wave/phase routing | specialized materials, noise, readout, integration maturity |
+| CMOS/GPU | controller, sensing, optimization, and high-throughput arithmetic | does not itself test the physical-field hypothesis |
+
+For the user's "loosen and densify" actuator, magnetorheological material is the closer
+engineering category. For mobile information packets, ferrofluid droplets are closer. For an
+eventual chip, spintronic films are the relevant scale.
+
+## Buildable hybrid prototype
+
+The first physical machine should be a peripheral coprocessor, not a replacement PC.
+
+```mermaid
+flowchart LR
+    G[Verified goal and policy] --> C[Bounded controller]
+    C --> D[Coil or electrode drivers]
+    D --> P[Magnetic pocket array]
+    P --> S[Camera, Hall, resistance, or RF readout]
+    S --> E[State estimator]
+    E --> C
+    E --> R[Signed transition receipt]
+    R --> G
+```
+
+The host PC performs policy, optimization, logging, and training. The physical array supplies
+nonlinearity, hysteresis, memory, routing, or actuation. This division lets the idea be tested
+without pretending a fluid board can match a GPU.
+
+## Acceptance ladder
+
+1. **Simulation:** predict each transition, include hysteresis/noise, and reproduce NAND, NOR,
+   XOR-through-depth, reset, and HOLD states. Compare against a size-matched digital control.
+2. **Macroscopic pocket board:** demonstrate closed-loop routing for at least 1,000 transitions;
+   publish transition error, settling time, energy, temperature, and cross-talk. Repeat across
+   at least three runs and devices.
+3. **Computation:** demonstrate universal gates, fanout of at least two, state retention, reset,
+   and cascaded operation. A successful gate in isolation is insufficient.
+4. **Reservoir task:** use a fixed physical reservoir plus trained linear readout. Beat both a
+   no-reservoir baseline and a parameter-matched software control by more than two pooled
+   standard deviations; otherwise report `UNDERPOWERED` or `NO_LIFT`.
+5. **Chip path:** supply a device model, SPICE or micromagnetic co-simulation, clock network,
+   I/O circuit, PDK-compatible layout, timing/power analysis, and fabrication partner.
+6. **Wafer performance:** only claim an accelerator after a compiled kernel beats the matched
+   dense/sparse baseline in wall time, energy, memory traffic, and task quality on the same
+   hardware.
+
+## Wafer relationship to current NVIDIA systems
+
+The shared problem is routing useful work and data among many compute units. NVIDIA's current
+platforms solve this through extreme codesign of GPUs, CPUs, switches, networking, memory, and
+software. The Clay wafer's current 64-cell/top-4 expert router is a software model of task
+placement. A graph coordinate is not a physical wire, six readout views are not six processors,
+and signed trits do not reduce storage unless an implemented encoding and kernel prove it.
+
+The near-term research claim is therefore:
+
+> SCBE can use a governed control language to address a simulated or tabletop magnetic state
+> array, while the wafer router can be evaluated as a sparse software scheduler. Physical
+> acceleration remains an experimental hypothesis.
+
+## Primary sources
+
+- Katsikis, Cybulski, and Prakash, [Synchronous universal droplet logic and control](https://www.nature.com/articles/nphys3341), *Nature Physics* 11 (2015).
+- Crepaldi et al., [Experimental Demonstration of In-Memory Computing in a Ferrofluid System](https://pubmed.ncbi.nlm.nih.gov/36919899/), *Advanced Materials* 35 (2023), DOI 10.1002/adma.202211406.
+- Colci et al., [Sub-nanosecond signal propagation in anisotropy-engineered nanomagnetic logic chains](https://www.nature.com/articles/ncomms7466), *Nature Communications* 6 (2015).
+- Currivan-Incorvia et al., [Logic circuit prototypes for three-terminal magnetic tunnel junctions with mobile domain walls](https://www.nature.com/articles/ncomms10275), *Nature Communications* 7 (2016).
+- Beneke et al., [Gesture recognition with Brownian reservoir computing using geometrically confined skyrmion dynamics](https://www.nature.com/articles/s41467-024-52345-y), *Nature Communications* 15 (2024).
+- NVIDIA, [Vera Rubin platform production announcement](https://nvidianews.nvidia.com/news/vera-rubin-full-production-agentic-ai-factory) (2026).
+
+This specification records engineering connections and disconfirming tests. It is not evidence
+that the integrated device has been built and is not a patent-novelty determination.

package.json

@@ -187,6 +187,7 @@
     "benchmark:swe-verified:readiness": "python scripts/benchmark/swe_verified_readiness.py",
     "benchmark:cli": "python scripts/benchmark/cli_competitive_benchmark.py",
     "benchmark:fleet-composition": "python scripts/benchmark/fleet_composition_stress.py",
+    "benchmark:fleet-transport": "python scripts/benchmark/bijective_frequency_transport.py",
     "aetherdesk:chemistry": "python scripts/aetherdesk_tile_actions.py chemistry",
     "aetherdesk:token": "python scripts/aetherdesk_tile_actions.py token",
     "aetherdesk:instrument": "python scripts/aetherdesk_tile_actions.py instrument",