# Learned latch placement and transfer V1: initialization improved, arithmetic and planning favored

Four independent feedback searches learned actual latch placements that reduced mean cost on untouched larger-host/capacity contexts relative to the paired boundary initialization: **8.40%, 15.00%, 2.46%, and 14.74%**. The four descriptive means span 2.46–15.00%, with equal-search mean 10.15%. Every selected policy nevertheless had higher mean cost than the simple analytical local rule and workload-aware conventional planning. Only searches 1 and 3 changed placement when workload cues changed. This is a learned structural decision, but no developmental comparative utility or necessity of learning was demonstrated.

All 512 candidate vectors, 8,192 training placements, 128 validation placements and 2,304 final conditions completed. All final conditions passed ordinary correctness, all-address/overwrite/reset and opposite-answer diagnostics, and compiled replay. There were zero final or candidate construction failures. Correctness is a gate under a common physical safety bound, not an accuracy contest. The full campaign began 2026-10-14T22:17:49.720098+00:00 and ended 2026-10-14T22:19:46.784538+00:00; publication status and receipt are separately maintained.

## Question and scope

Can one compact shared local placement scorer trained by aggregate feedback produce lower-cost organization that transfers to fresh hosts and L6/K8 beyond its competent initialization and ordinary local arithmetic, and how does it compare with workload-aware global planning? Only the scalar latch's site on an available write-to-query tree path is learned. Fixed local exploration, tree topology, task algorithm, primitives and scheduler are supplied. There is no online task learning, within-episode development, inherited learned-state growth or discovery of task memory. The broader DNA-like developmental hypothesis remains open.

The earlier de-scaffolded constructor-only report, handoff, configuration, actual source and closeout were read from completed local artifacts after the supplied public URL was inaccessible through the web tool. That milestone is closed. D stored at write attachments; B independently planned weighted shortest paths and placed latches nearer the query. Its former 1292-tick envelope was a correctness bound, not an optimization target. This new task/objective is versioned V3; old results are preserved and never substituted for the new workload costs.

## Primary transfer effects and all controls

The prespecified primary endpoint equally averages 144 paired host/profile contexts per policy: six settings with L6 or K8, two workloads, twelve groups. A positive percentage means (comparator J_star − learned J_star) / comparator J_star. References execute once per context and are shared across policies. Each search is reported; no best-of-four selection occurs.

| Search | Versus D_initial | Versus D_analytic | Versus B_original | Versus B_cost |
| --- | ---: | ---: | ---: | ---: |
| 0 | +8.40% | -12.57% | -0.09% | -14.57% |
| 1 | +15.00% | -5.56% | +5.20% | -7.55% |
| 2 | +2.46% | -25.49% | -14.59% | -28.15% |
| 3 | +14.74% | -5.72% | +5.11% | -7.71% |

Raw paired tick reductions use the same weighting; the final column is learned cost minus the exact feasible D-tree minimum, so positive values are suboptimality gaps.

| Search | Initial minus learned | Analytic minus learned | Original B minus learned | Cost B minus learned | D-tree optimum gap |
| --- | ---: | ---: | ---: | ---: | ---: |
| 0 | +2.501 | -0.791 | -0.016 | -0.998 | 0.982 |
| 1 | +2.947 | -0.345 | +0.430 | -0.552 | 0.536 |
| 2 | +0.699 | -2.593 | -1.818 | -2.800 | 2.784 |
| 3 | +2.949 | -0.343 | +0.432 | -0.551 | 0.534 |

B_original was designed for query proximity; two policies improve its new workload cost by about 5%, while search 0 approximately matches and search 2 loses. That comparison cannot establish utility: B_cost is the competent workload-aware control and all four lose to it on mean transfer cost. B uses its own independent tree. D-tree optimality is only over eligible distinct assignments on D's fixed tree, not all trees or runtimes. No learned result beats that bound.

The fresh L4/K4 references are separate from size transfer. Initial-versus-learned effects there are 9.19%, 12.48%, 2.99%, 13.53% for searches 0..3. Across all 192 contexts the effects are 8.60%, 14.37%, 2.59%, 14.44%. L6 and K8 were historically studied, but absent from this new policy optimizer; they are new policy-transfer dimensions, not historically unknown dimensions.

## Transfer by group, setting and workload

Each group averages its twelve primary contexts. All four policies improve every group mean, but individual contexts can worsen: the worst initial contrast is −40.76% for search 0. Group means and thousands of transactions are not independent algorithm discoveries. No significance hunting, precise broad confidence interval or powered-equivalence claim is made from four searches.

| Group | Search0 | Search1 | Search2 | Search3 |
| --- | ---: | ---: | ---: | ---: |
| 0 | 9.59% | 14.88% | 0.69% | 15.68% |
| 1 | 8.92% | 16.22% | 1.88% | 17.64% |
| 2 | 6.51% | 15.04% | 3.48% | 14.17% |
| 3 | 10.48% | 15.46% | 1.85% | 14.42% |
| 4 | 9.27% | 14.00% | 3.88% | 13.33% |
| 5 | 7.66% | 15.96% | 3.14% | 14.73% |
| 6 | 9.02% | 15.99% | 2.63% | 17.69% |
| 7 | 8.76% | 15.34% | 2.76% | 14.65% |
| 8 | 7.06% | 12.80% | 1.72% | 12.29% |
| 9 | 9.66% | 16.54% | 2.26% | 17.16% |
| 10 | 7.73% | 14.12% | 1.61% | 12.94% |
| 11 | 6.15% | 13.69% | 3.60% | 12.19% |

The complete fixed stratification below equally averages twelve groups within each setting/profile. It shows how a positive mixed-workload mean can coexist with worse write-heavy conditions.

| Host/profile | Search0 vs initial | Search1 vs initial | Search2 vs initial | Search3 vs initial |
| --- | ---: | ---: | ---: | ---: |
| L4 K4 p0, rho 0.25 | -12.44% | -5.50% | -6.05% | -2.75% |
| L4 K4 p0, rho 0.75 | +32.83% | +32.71% | +11.53% | +32.83% |
| L4 K4 p0.15, rho 0.25 | -12.75% | -6.60% | -6.56% | -5.09% |
| L4 K4 p0.15, rho 0.75 | +29.13% | +29.32% | +13.04% | +29.13% |
| L4 K8 p0, rho 0.25 | -8.36% | -3.57% | -7.07% | -3.97% |
| L4 K8 p0, rho 0.75 | +28.12% | +28.12% | +13.94% | +28.12% |
| L4 K8 p0.15, rho 0.25 | -12.29% | -3.78% | -6.50% | -6.18% |
| L4 K8 p0.15, rho 0.75 | +24.71% | +24.64% | +12.60% | +24.71% |
| L6 K4 p0, rho 0.25 | -22.67% | -6.29% | -5.27% | -4.46% |
| L6 K4 p0, rho 0.75 | +36.93% | +36.35% | +10.28% | +36.93% |
| L6 K4 p0.15, rho 0.25 | -20.55% | -2.27% | -4.35% | +0.49% |
| L6 K4 p0.15, rho 0.75 | +40.18% | +39.39% | +8.45% | +40.18% |
| L6 K8 p0, rho 0.25 | -20.40% | -2.18% | -4.70% | -6.52% |
| L6 K8 p0, rho 0.75 | +36.30% | +36.12% | +8.91% | +36.30% |
| L6 K8 p0.15, rho 0.25 | -17.88% | -3.14% | -5.29% | -5.43% |
| L6 K8 p0.15, rho 0.75 | +36.72% | +36.66% | +8.50% | +36.72% |

## Causal workload-cue mismatch

On the same geometry, importance draws, events and true objective, each selected scorer also constructs with the other rho profile's cues. The table reports primary mismatch minus intact cost. The same coefficients must respond through the local construction program; no future workload inference or online adaptation is claimed.

| Search | Host pairs whose two profiles change placement | Mismatch increase, ticks | Mean fractional increase | Increased / equal / decreased primary contexts |
| --- | ---: | ---: | ---: | ---: |
| 0 | 0/96 | +0.000 | +0.00% | 0 / 144 / 0 |
| 1 | 95/96 | +1.898 | +18.11% | 133 / 2 / 9 |
| 2 | 0/96 | +0.000 | +0.00% | 0 / 144 / 0 |
| 3 | 87/96 | +1.240 | +13.07% | 127 / 12 / 5 |

All four learned policies differ from initialization in all 192 contexts. Searches0 and 2 have identical placements under both cues: mismatch changes neither placement nor true cost. Searches1 and 3 show substantial context dependence, with occasional beneficial mismatches retained rather than excluded. Their mismatched costs rise on average by 18.11% and 13.07%. This is partial evidence for learned context-sensitive placement, with two informative failures of cue sensitivity.

## Runtime reduction, local mechanism and exact controls

For address a on a symmetric rooted tree, let D_a be write depth, d_a latch depth and R response depth, in physical transport ticks. Atomic router/latch processing adds zero delay. Writes require **1+D_a−d_a** ticks from external injection to the unique intended latch commit. Queries require **2+2d_a+R** ticks to the single correct external response. The one-tick terminal links, real internal bond delays, interface handshake and response path are included. The evaluator observes commits; the executing system receives no free acknowledgment. Therefore

    J = sum_a [w_a(1+D_a−d_a) + q_a(2+2d_a+R)].

Idle settlement gaps, initialization and fixed diagnostics do not enter stationary J. The variable local term is w*x+2q*d. Positive per-address importance draws in {1,2,4} normalize separately for writes/reads and scale by 1−rho/rho. Both rho=.25/.75 profiles share importance draws; all methods receive the same public description through the common construction adapter. The unchanged B_original algorithm ignores workload weights; B_cost uses them. Event signs and address importance are unrelated to favored sites. Samples have 64 subsequent events per episode, with no forced query minimum. Across the 192 shared ordinary episodes per workload profile, the realized read fractions are .25016 and .74797, rather than exactly .25/.75.

The previous positive-delay host generator and executable growth program are unchanged. Charged births and registrations discover a fixed tree. A local demand token travels query→write to carry q and accumulate path delay; each write attachment supplies its own w. Returning bids visit eligible vacant sites using raw local x/32, d/32, 8w and 8q features. A four-unit tanh scorer has 25 bounded binary64 coefficients, including a disclosed own-write prior. The root receives only the reduced winning score/handle, not the host or a cost matrix. The local reserve token differentiates the chosen latch and installs executable Q/W routes; an acknowledgment completes before the next address. The scheduler delivers/enforces/rejects and serializes; it never assigns or repairs a latch. Addresses are ordered ascending; score ties use opaque handles. Root, response and other write attachments are excluded, distinct reservations are required, and the own write site is always feasible.

THETA0 has fixed authored hidden weights, zero output weights and own-write coefficient −.25. It faithfully reproduces the prior boundary placement through the same mechanism. It is a competent hand-authored initialization, not an artificially broken random baseline. D_analytic uses immediate arithmetic in this identical shell. B_original's builder source remains exact previous bytes except its imported compatible shared runtime; it ignores workloads as before. B_cost retains independent weighted Dijkstra/ties and solves the new distinct-site assignment exactly with integer costs. The evaluator-only D-tree assignment is separate from learning and unavailable during search. No optimal labels, precomputed matrix, objective products as features, test normalization or whole-host root gathering enter the policy.

Tiny exhaustive allocations and twenty integer matrices validated the Hungarian solver before optimization. Post-run independent saved-data analysis reconstructs both trees from saved geometry, verifies every placement's timing/eligibility and checks both exact bounds using a separate subset dynamic program. All other deployment limits are nonbinding across this assignment space: <=216 cells/215 tree edges, finite route keys, conservatively <=3870 route entries within 1 MiB, one sequential request/value/response transaction with <=647 messages and one live queue record, and bounded terminal/output buffers. The independent 1292-tick T_cap derives from three paths of at most 215 two-tick edges plus two terminal ticks; writes are <=431. Gaps are 1294..1394; episode cap 110000 covers <=72 events. Identifiers recycle within 0..63 only after full settlement/output drain. Invalid models would receive J_star=1292 as a penalty, separately from raw measured cost; aggregate-unexecuted conditions would remain unobserved. No such final failures occurred.

## Training, validation and two seals

Each search uses eight fresh L4/K4 training hosts and four independent validation hosts, balanced p0/.15 with both workloads. Runs and development/training/validation/test use disjoint namespaces. Every candidate reruns the actual local placement shell and short functional diagnostics on all sixteen training contexts; feedback is its own correctness-gated expected cost. Seeded Gaussian mutation perturbs all coordinates and clips bounds; strict training improvements update the incumbent, earlier indexes win ties. Scales/budgets/repeat handling are in config.json. Exactly 128 vectors/run include theta0. No second optimizer, manual post-score coefficient edit or candidate-budget increase occurred.

Three best distinct training vectors plus theta0 undergo the prespecified validation/padding rule. Validation selects by mean cost then earlier index and feeds no further mutation. Initial and selected train/validation scores remain separate from test scores.

| Search | Selected evaluation | Training initial → selected J_star | Validation initial → selected J_star |
| --- | ---: | ---: | ---: |
| 0 | 11 | 8.440 → 7.204 | 9.358 → 7.779 |
| 1 | 112 | 9.201 → 7.128 | 8.594 → 6.661 |
| 2 | 4 | 8.558 → 7.838 | 10.669 → 10.202 |
| 3 | 62 | 9.225 → 7.540 | 9.606 → 7.331 |

Optimization freeze SHA256 **0c1e93875f8e5a7abc182c22412a4fbb199519e337257c38a030e280662c4178** binds executable source, parameterization, objective, optimizer, splits, stopping, resources and development fixtures before optimization. It contains no learned coefficients. Selection seal SHA256 **9860e0a246b3f067d57c7dbcfb4a2fc0daa2a068e674fd993cef79d4de8e725d**, at 2026-10-14T22:18:09.033519+00:00, binds all four selected policies, initial controls, full lineage/training/validation records and unchanged source before any test outcomes are generated. Frozen source remains unchanged. Every parent, accepted/rejected vector, context score and candidate diagnostic is saved.

Fresh test settings are independently generated within each outer group using setting-bearing keys. Profiles share each host and importance draws; all methods share the same external event streams. Four policies and twelve groups cross each other while sharing a task family. Four policy searches are the learning replications; the four shared controls are not executed or counted four times.

## Expected versus sampled cost, state and traffic

Each method has 192 final contexts. J below is the equally weighted exact expected cost over all settings; the sampled mean is the mean actual service-completion latency of the 128 subsequent events/context. They differ by design. Initialization and fixed diagnostics are separately costed below. All ordinary 147,180 queries and 147,732 sampled writes passed; 27,648 initialization writes also committed correctly. All 50,688 diagnostic queries and 55,296 diagnostic/setup writes passed, including 4,608 opposite-answer pairs. All 4,608 ordinary episode replays passed, comparing 3,410,527 delivery/boundary state steps.

| Method | Exact J | Realized sampled mean | Mean active cells | Mean persistent field bytes | Stationary messages |
| --- | ---: | ---: | ---: | ---: | ---: |
| D_initial | 11.373 | 11.331 | 28.06 | 283.72 | 232,973 |
| D_learned_0 | 9.066 | 9.042 | 28.06 | 283.72 | 186,432 |
| D_learned_1 | 8.687 | 8.665 | 28.06 | 283.72 | 178,498 |
| D_learned_2 | 10.693 | 10.652 | 28.06 | 283.72 | 220,404 |
| D_learned_3 | 8.673 | 8.641 | 28.06 | 283.72 | 178,018 |
| D_analytic | 8.360 | 8.322 | 28.06 | 283.72 | 171,853 |
| B_original | 9.053 | 9.020 | 27.78 | 219.84 | 185,646 |
| B_cost | 8.181 | 8.151 | 27.78 | 238.20 | 168,744 |
| D_mismatch_0 | 9.066 | 9.042 | 28.06 | 283.72 | 186,432 |
| D_mismatch_1 | 10.444 | 10.413 | 28.06 | 283.72 | 215,674 |
| D_mismatch_2 | 10.693 | 10.652 | 28.06 | 283.72 | 220,404 |
| D_mismatch_3 | 9.797 | 9.801 | 28.06 | 283.72 | 202,613 |

Mean D wire length is 26.475 and mean route entries 64.531 for every D policy; B wire 26.246, with B_original 43.427 and B_cost 49.547 route entries. D retains the fixed terminal path union, so moving latches changes traffic/latency without reducing this shell's cell/route allocation. D's mean temporary occupancy is 137.25 sites versus B 27.78. D development takes mean 682.82 messages/21,235.17 logical bytes and 178.49 ticks/deployment, including charged four-pass placement. These costs are identical across theta/analytical choices because the shell always traverses each full write path. B sends no development transport but globally plans: mean work units 1,048.66 for B_original and 1,223.67 for B_cost, including actual assignment work. These different work units are not FLOPs or simulated ticks.

Every executing circuit reserves 8,324 compact bytes for finite queue/interface/output buffers; observed queue peak is 1 and buffered field peak 28 bytes. Observed full serialized mutable state peaks at 229 bytes. These field/serialized quantities are not exact Python object memory. Scorer coefficients are 200 binary64 bytes each (JSON sizes separately listed); the actual growth+placement program is 6,049 UTF-8 bytes, scheduler 6,308, runtime 7,430, in addition to host, task, primitive, metadata and optimizer sources. Parameter count does not imply entire-program information compression. A shared immutable instruction bundle/coefficients is supplied to local updates; its bytes are separate from charged task/development message payload.

All stationary messages carry four payload bytes; replay repeats ordinary initialization+sample work and is recorded separately. Costs below are total service ticks/messages across 192 contexts, not additions to J. Fixed diagnostics include all-address overwrites, reset zeros, two cloned pairs and their setup.

| Method | Initialization ticks / messages | Diagnostics ticks / messages | Replay messages |
| --- | ---: | ---: | ---: |
| D_initial | 2,304 / 2,304 | 96,624 / 80,764 | 235,277 |
| D_learned_0 | 12,392 / 10,624 | 78,672 / 65,964 | 197,056 |
| D_learned_1 | 9,834 / 8,432 | 83,266 / 69,896 | 186,930 |
| D_learned_2 | 5,512 / 4,724 | 91,296 / 76,736 | 225,128 |
| D_learned_3 | 11,080 / 9,492 | 81,036 / 68,004 | 187,510 |
| D_analytic | 9,102 / 7,894 | 84,302 / 70,614 | 179,747 |
| B_original | 12,464 / 10,560 | 78,568 / 65,772 | 196,206 |
| B_cost | 9,174 / 7,854 | 84,622 / 70,738 | 176,598 |
| D_mismatch_0 | 12,392 / 10,624 | 78,672 / 65,964 | 197,056 |
| D_mismatch_1 | 9,834 / 8,432 | 83,266 / 69,896 | 224,106 |
| D_mismatch_2 | 5,512 / 4,724 | 91,296 / 76,736 | 225,128 |
| D_mismatch_3 | 11,080 / 9,492 | 81,036 / 68,004 | 212,105 |

## Historical selection cost and full ledger

The complete monitored campaign, including candidate construction/feedback, validation, final deployment, all baselines/oracles and replay, used **131.887 CPU / 117.218 executed-job wall seconds**, including monitor overhead. Current study jobs plus conservative failed-start reservation use 141.788 CPU/126.510 wall seconds. A separate 120-second setup/analysis/publication-bookkeeping allowance is retained. Inherited inclusive costs remain 444.820 CPU/410.215 wall seconds; they include earlier failed-monitor reservations and were checked against saved ledgers. Conservative combined totals are **717.284 CPU / 667.415 wall seconds**, below 3600/5400 ceilings. Remote publication CPU is unmeasured. The separate bounded static packing check adds 10.675 CPU/10.690 wall seconds to these conservative totals; its OS-reported process RSS peak was 630,505,472 bytes, below 1 GiB. It changes transport encoding only; scientific source, outcomes and first-sealed hashes are unchanged.

The resource artifact records every job and failed attempt, 512 vectors, 8,192 training and 128 validation constructions, 98,304/1,536 short feedback transactions, 1,920 final D placements and 384 B plans, 192 separate D-tree assignments,322,560 ordinary initialization+sample transactions and 322,560 replay transactions, 105,984 fixed diagnostic transactions, and the prescribed 12-condition author rerun. Growth is actually discovered once for 48 training/validation and 96 test hosts and reused as pure local state; every D placement reruns the shell. Per-deployment costs include the growth charge; actual numerical work counts each cached discovery once. Validation padding alone may cache a duplicate theta0 slot; no such extra candidate evaluation occurs. Full monitored CPU includes shared serialization/validation/objective work omitted by narrow stage timers.

Highest sampled process-tree RSS was **313,999,360 bytes (299.45 MiB)**. CPU/file-size rlimits and the inherited 100ms RSS/wall/artifact watchdog enforce finite jobs; sampled RSS is not an instantaneous hard guarantee. The first sandbox-denied monitor preflight launched no scientific worker, is preserved locally and charged one second. No GPU, scheduler, extra agent, service, numerical framework, global install or measured-model inference was used. Scoped static evidence is losslessly packed into supported JSON rather than duplicating large pretty-printed records, keeping study+release below 128 MiB; final exact byte count is in local closeout.

No defensible amortization break-even is claimed. Historical selection adds CPU/feedback work, D per-instance construction remains around 3.67–3.70 ms inclusive of cached-growth timing attribution, versus B_cost 0.50 ms per plan, and all selected policies lose B_cost on primary expected service cost. Deployments and transactions/deployment do not rescue that mean latency comparison. Python CPU milliseconds, simulated ticks and communication bytes cannot be equated. B_cost's actual per-instance planning cost is in every record.

## Strongest counterexample and separate verdicts

Prespecified maximum adverse learned-versus-B_cost case: **search 2, group0/L6/K4/p0/rho0.75**. Learned J=23.050, B_cost=11.533, D_analytic=11.750, D-tree optimum=11.317, initial=24.417. All pass correctness. The learned circuit improves slightly over initialization while costing 11.517 more ticks than B_cost—almost twice its expected cost. This directly defeats an inference from learned improvement to comparative utility.

1. **Learning:** supported for this narrow structural decision. Feedback changes placement and all four selected policies improve held-out group means beyond a competent boundary initial policy, with unequal individual-context effects.
2. **Transfer and context:** average benefit survives the specified L/K transfer; workload-cue dependence appears only in searches 1/3 and fails in 0/2. This is context-sensitive construction in two selected policies, not unknown-future-workload inference or online adaptation.
3. **Necessity and utility:** unsupported. Ordinary local arithmetic beats every policy's primary mean, and workload-aware conventional planning also wins with lower construction work/temporary occupancy. Learning can approximate a cheap structural solution here but has not made it necessary or competitively useful.

An analytical control outperforming learning is a substantive negative comparative result. No task/search/criterion was changed to rescue the claim. This treatment is seeded feedback parameter search; it was not compared with random search, gradients or alternative learning algorithms. Exact solver success, 25 parameters and perfect correctness are not the scientific endpoint.

## Evidence, checks and disposable commands

Source/configuration, both seals, selected coefficients, every candidate/selection record, full host/workload/event manifests, all final outcomes/costs, exact bounds and prespecified full construction/delivery/state trajectories are delivered. Large evidence is stored as lossless tar/XZ/base64 in a static evidence-pack.json with decoded byte counts and SHA256s. unpack_evidence.py restores exact sealed originals without executing content. This uses existing allowed JSON/text publication formats, with no publisher/policy modification. Raw private chats, unrelated logs, whole repositories and separate-primary state are excluded. Old originals/releases remain intact; scoped preservation hashes are provenance, not fresh evidence verification for a standalone downloader.

The saved-data author audit recomputes all external-history targets/gates, obtains J from observed write commits/all-address query completions, verifies optimizer lineage/validation selection, reconstructs trees independently and cross-checks exact optima with subset dynamic programming. It imports no experimental modules. The prescribed first selected policy on group0/L6/K8/p.15, both workloads, controls and mismatch matched compiled hashes, observations, trajectories and costs in a fresh disposable interpreter: twelve conditions. These are author checks, not independent reproduction. Compiled replay starts with only generated executable structure/parameters and initialized execution state; no development history or training store enters it.

Restore the exact selected bundle:

```sh
python3 unpack_evidence.py --out /tmp/latch-evidence
```

Within that restored directory, one development-only smoke command and one bounded full-campaign command create fresh disposable directories and carry the accumulated ledger:

```sh
python3 reproduce.py smoke --out /tmp/latch-smoke
python3 reproduce.py full --out /tmp/latch-campaign
```

Each output must not already exist. The closed original directory refuses duplicate campaigns. To audit saved records without rerunning the experiment, run python3 run_job.py audit in the restored directory. The complete source freeze and artifacts—not just policy.py—are required. The full reproduction command is supplied, not executed as a second campaign here.

Content inspection, visual inspection, saved-data audit, executable author rerun and publication verification are distinct records. Publication is routine authorized delivery, neither scientific replication nor peer review. A concise editorial note about the stale Research overview was sent to the user-designated publisher-owner chat; summary correction is separate and is not a publication prerequisite. No shared-site edit is requested here.

At most one next scientific action is recommended: a separately bounded study of competing local allocation where simple immediate arithmetic is no longer sufficient, retaining its analytical and workload-aware planning controls. It is not executed.