jgrusewski 08fd5803c4 refactor(sp22): H6 Phase 3 α cleanup + runbook revision for atom-shift design ...

Same-session cleanup of Phase 3b scalar-bias residue (commit cb80b74ce's
additions that became architecturally redundant under the 2026-05-13
atom-shift revision). The scalar-bias α design is mathematically
ineffective in C51 distributional Q-learning (softmax-shift-invariance
→ W gradient = 0 → never trains). The revised design threads
`aux_atom_shift[b, a] = W[a] * state_121[b]` through compute_expected_q
+ c51_loss_kernel + c51_grad_kernel + mag_concat_qdir; dW + dstate
gradients integrate into c51_grad_kernel's projection backward.

Files
─────
- crates/ml/build.rs:
    Removed `aux_to_q_dir_bias_kernel.cu` + `aux_to_q_dir_bias_backward_kernel.cu`
    from kernels_with_common. Replaced with comment block documenting
    C51 softmax-invariance reason + spec/audit pointers.

- crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs:
    Removed `pub(crate) static SP22_AUX_TO_Q_DIR_BIAS_CUBIN` and
    `SP22_AUX_TO_Q_DIR_BIAS_BWD_CUBIN` static declarations.
    Removed 3 struct fields (aux_to_q_dir_bias_kernel,
    aux_to_q_dir_bias_backward_dw_kernel,
    aux_to_q_dir_bias_backward_dstate_kernel) and their new()
    loading blocks + struct construction entries.
    KEPT: w_aux_to_q_dir + adam_m_w_aux + adam_v_w_aux + dw_aux_buf
    (still needed for atom-shift Adam-trained W).
    Updated W doc-comment to describe atom-shift threading (was
    scalar-bias) and structural-prior initialization `[-0.5, 0.0,
    +0.5, 0.0]`.

- crates/ml/src/cuda_pipeline/aux_to_q_dir_bias_kernel.cu +
  crates/ml/src/cuda_pipeline/aux_to_q_dir_bias_backward_kernel.cu:
    Source files stay on disk as committed dead code (commit
    464bc5f7a history preserved). nvcc no longer compiles them
    (build.rs unregistered). No `include_bytes!` references remain.

- docs/plans/2026-05-13-sp22-h6-phase3-alpha-beta-runbook.md:
    Tasks A3/A4/A5 marked SUPERSEDED; A5 retains cleanup instructions.
    Task B9 Steps 4-10 rewritten as Steps 4 (cleanup), 5 (W structural
    prior init), 6 (compute_expected_q atom-shift threading),
    7 (c51_loss_kernel), 8 (c51_grad_kernel backward dW + dstate),
    9 (mag_concat_qdir), 10 (quantile_q_select/iqn_dual_head
    investigation), 11 (Adam wireup), 12 (verify + capture).
    Task C1 rewritten: collector passes 4 more args through
    existing compute_expected_q launcher (NO new kernel).

- docs/dqn-wire-up-audit.md:
    Cleanup-commit entry documenting the runbook revision and the
    code-cleanup actions.

Verification
────────────
- cargo check -p ml --features cuda: 0 errors, 21 pre-existing
  warnings (Phase 3b baseline parity).
- Build script no longer compiles the deleted-registration kernels.

Phase 3-final scope (~40-60 hr engineering — unchanged)
───────────────────────────────────────────────────────
Implementation per the revised runbook: atom-shift threading
through 4-5 kernels (compute_expected_q, c51_loss_kernel,
c51_grad_kernel, mag_concat_qdir, possibly quantile_q_select/
iqn_dual_head) + Adam wireup + collector launcher arg passing +
A2 eval-side + SP11 controller extension + HEALTH_DIAG telemetry
+ verification gates + atomic Phase F commit + smoke + verdict.

Refs
────
- docs/plans/2026-05-12-sp22-h6-phase3-alpha-beta.md (spec α section
  revised 2026-05-13 — commit 648078ce2)
- docs/plans/2026-05-13-sp22-h6-phase3-alpha-beta-runbook.md (runbook
  α tasks revised in this commit)
- 464bc5f7a (Phase A — α kernels added; now committed dead code)
- cb80b74ce (Phase 3b — α struct fields + cubin statics added; cleaned
  up in this commit)
- pearl_no_partial_refactor (atom-shift threading is the new atomic
  contract for direction-branch atom positions)

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

2026-05-13 01:24:15 +02:00

..

2026-02-26-on-demand-training-design.md

docs: add on-demand training dispatch design

2026-02-26 11:39:05 +01:00

2026-02-26-on-demand-training-implementation.md

docs: add on-demand training dispatch implementation plan

2026-02-26 11:42:54 +01:00

2026-02-26-training-pipeline-design.md

feat(ml): add manually-triggered GitLab CI training pipeline

2026-02-26 09:04:58 +01:00

2026-02-26-training-pipeline-implementation.md

feat(ml): add manually-triggered GitLab CI training pipeline

2026-02-26 09:04:58 +01:00

2026-02-27-autonomous-agents-design.md

docs: autonomous CI agents design — OpenHands + GitLab issue bus

2026-02-27 08:14:17 +01:00

2026-02-27-autonomous-agents-implementation.md

docs: add autonomous agents implementation plan (23 tasks, 6 phases)

2026-02-27 08:20:56 +01:00

2026-02-27-cockpit-nat-gateway-design.md

docs: add Cockpit metrics + NAT gateway design

2026-02-27 00:27:23 +01:00

2026-02-27-cockpit-nat-gateway-implementation.md

docs: add implementation plan for Cockpit metrics + NAT gateway

2026-02-27 00:30:06 +01:00

2026-02-27-gpu-resource-optimization-design.md

docs: GPU resource optimization design — RL→L4, supervised→L40S

2026-02-27 16:15:07 +01:00

2026-02-27-gpu-resource-optimization-implementation.md

docs: GPU resource optimization implementation plan

2026-02-27 16:16:55 +01:00

2026-02-27-training-deploy-design.md

docs: add training pipeline & model deployment design

2026-02-27 22:08:26 +01:00

2026-02-27-training-deploy-plan.md

docs: add training pipeline & model deployment implementation plan

2026-02-27 22:13:29 +01:00

2026-02-28-gpu-pipeline-design.md

perf: eliminate heap-alloc Decimal→float casts across 19 files (36 instances)

2026-02-28 02:29:04 +01:00

2026-02-28-gpu-pipeline-phase2b-design.md

docs: add GPU pipeline Phase 2b design — zero-roundtrip CUDA experience collection

2026-02-28 11:51:43 +01:00

2026-02-28-gpu-pipeline-phase2b-implementation.md

docs: add Phase 2b implementation plan — 10 tasks for zero-roundtrip kernel

2026-02-28 11:58:30 +01:00

2026-02-28-gpu-pipeline-phase2c-implementation.md

docs: add PPO CUDA pipeline Phase 2c implementation plan — 11 tasks

2026-02-28 13:07:20 +01:00

2026-02-28-gpu-utilization-optimization-design.md

docs: GPU utilization optimization design

2026-02-28 21:18:51 +01:00

2026-02-28-gpu-utilization-optimization-implementation.md

docs: GPU utilization optimization implementation plan

2026-02-28 21:24:51 +01:00

2026-02-28-numeric-type-standardization-design.md

docs: numeric type standardization design (i64/6dp financial types)

2026-02-28 02:32:46 +01:00

2026-02-28-numeric-type-standardization-implementation.md

docs: numeric type standardization implementation plan

2026-02-28 02:37:53 +01:00

2026-02-28-ppo-cuda-pipeline-design.md

docs: add PPO CUDA pipeline Phase 2c design — zero-roundtrip experience collection

2026-02-28 12:59:45 +01:00

2026-02-28-s3-binary-share-design.md

docs: add S3 binary share design — eliminate Docker builds, faster pod startup

2026-02-28 17:46:13 +01:00

2026-02-28-s3-binary-share-implementation.md

docs: add S3 binary share implementation plan (15 tasks)

2026-02-28 17:50:37 +01:00

2026-03-01-codebase-consolidation-design.md

docs: add codebase consolidation & DRY cleanup design

2026-03-01 19:22:22 +01:00

2026-03-01-codebase-consolidation-implementation.md

docs: add codebase consolidation implementation plan

2026-03-01 19:25:01 +01:00

2026-03-01-deferred-stubs-design.md

docs: add deferred stubs implementation design

2026-03-01 21:17:30 +01:00

2026-03-01-deferred-stubs-implementation.md

docs: add deferred stubs implementation plan

2026-03-01 21:21:53 +01:00

2026-03-01-gpu-blockers-validation-design.md

feat(ml): maximize GPU utilization — BF16 mixed precision, dynamic sizing, sync reduction

2026-03-01 17:54:25 +01:00

2026-03-01-gpu-full-sweep-design.md

docs: GPU optimization full sweep design — training binary, PPO parity, NCCL multi-GPU

2026-03-01 21:25:36 +01:00

2026-03-01-gpu-full-sweep-implementation.md

docs: GPU full sweep implementation plan — 17 tasks covering P0-P3

2026-03-01 21:25:36 +01:00

2026-03-01-gpu-max-performance-phase2-design.md

docs: GPU max performance Phase 2 design (L4 → H100)

2026-03-01 18:37:03 +01:00

2026-03-01-infra-security-hardening-design.md

docs: add infrastructure security hardening design

2026-03-01 23:13:13 +01:00

2026-03-01-infra-security-hardening-implementation.md

docs: add infrastructure security hardening implementation plan

2026-03-01 23:16:44 +01:00

2026-03-01-io-pipeline-wiring-design.md

docs: I/O pipeline wiring design — prefetcher, buffer pool, double-buffer

2026-03-01 21:39:50 +01:00

2026-03-01-io-pipeline-wiring-implementation.md

docs: I/O pipeline wiring implementation plan — 7 tasks

2026-03-01 21:42:38 +01:00

2026-03-01-ml-inference-production-cleanup-design.md

docs: ML inference production cleanup design

2026-03-01 17:16:08 +01:00

2026-03-01-ml-inference-production-cleanup-implementation.md

docs: ML inference production cleanup implementation plan

2026-03-01 17:23:07 +01:00

2026-03-01-real-metrics-overhaul-design.md

docs: add real metrics overhaul design

2026-03-01 01:51:24 +01:00

2026-03-01-real-metrics-overhaul-implementation.md

docs: add real metrics overhaul implementation plan

2026-03-01 02:00:08 +01:00

2026-03-01-training-metrics-common-design.md

docs: training metrics design — common::metrics ownership

2026-03-01 15:23:41 +01:00

2026-03-01-training-observability-design.md

docs: training observability + Grafana dashboards design

2026-03-01 22:25:11 +01:00

2026-03-01-training-observability-implementation.md

docs: training observability implementation plan (reduced scope)

2026-03-01 22:31:54 +01:00

2026-03-01-workspace-cleanup-design.md

docs: workspace cleanup design and implementation plan

2026-03-01 22:32:54 +01:00

2026-03-01-workspace-cleanup-implementation.md

docs: workspace cleanup design and implementation plan

2026-03-01 22:32:54 +01:00

2026-03-02-additional-training-metrics-design.md

docs: additional training metrics design — 17 new Prometheus metrics

2026-03-02 23:44:45 +01:00

2026-03-02-additional-training-metrics.md

docs: implementation plan for 17 additional training metrics

2026-03-02 23:48:18 +01:00

2026-03-02-binary-cache-consolidation-design.md

docs: binary cache PVC consolidation design (8 PVCs → 1 shared)

2026-03-02 10:58:33 +01:00

2026-03-02-binary-cache-consolidation-implementation.md

chore(infra): delete 8 per-service binary-cache PVCs (replaced by foxhunt-binaries)

2026-03-02 11:20:03 +01:00

2026-03-02-fxt-broker-check-design.md

docs: add design for fxt broker check CLI command

2026-03-02 13:26:17 +01:00

2026-03-02-fxt-broker-check-implementation.md

docs: add implementation plan for fxt broker check command

2026-03-02 13:28:51 +01:00

2026-03-02-gpu-per-sumtree-design.md

docs: GPU-resident PER sum-tree design

2026-03-02 23:40:57 +01:00

2026-03-02-gpu-per-sumtree-implementation.md

docs: GPU PER sum-tree implementation plan (13 tasks)

2026-03-02 23:45:27 +01:00

2026-03-02-monitoring-service-design.md

docs: add monitoring service design — live training metrics over gRPC

2026-03-02 21:05:48 +01:00

2026-03-02-monitoring-service.md

feat(fxt,ml): add fxt train monitor command and update DQN tests for action collapse fix

2026-03-04 22:58:40 +01:00

2026-03-02-node-pool-split-design.md

docs: 3-pool node split design (gitlab/platform/foxhunt)

2026-03-02 09:49:46 +01:00

2026-03-02-node-pool-split-implementation.md

infra(tf): add platform pool, rename services → foxhunt in Kapsule module

2026-03-02 09:59:21 +01:00

2026-03-03-bf16-training-design.md

docs: add BF16 training design — pure BF16 for all 10 models

2026-03-03 15:56:14 +01:00

2026-03-03-bf16-training.md

docs: update stale rainbow_agent_impl path in bf16 plan

2026-03-04 18:36:28 +01:00

2026-03-03-dqn-action-collapse-fix.md

docs: DQN action collapse fix design — 5 fixes for 45-action diversity

2026-03-03 00:59:20 +01:00

2026-03-03-dqn-production-hardening-design.md

docs: add DQN production hardening design

2026-03-03 20:40:46 +01:00

2026-03-03-epoch-financial-metrics-design.md

feat(fxt,ml): add fxt train monitor command and update DQN tests for action collapse fix

2026-03-04 22:58:40 +01:00

2026-03-03-epoch-financial-metrics-implementation.md

feat(fxt,ml): add fxt train monitor command and update DQN tests for action collapse fix

2026-03-04 22:58:40 +01:00

2026-03-03-fxt-dashboard-design.md

docs: fxt dashboard & CLI completion design

2026-03-03 01:14:02 +01:00

2026-03-03-fxt-dashboard-implementation.md

docs: fxt dashboard implementation plan (15 tasks)

2026-03-03 01:20:27 +01:00

2026-03-03-fxt-overhaul-design.md

docs: fxt 2.0 design and implementation plan

2026-03-03 20:11:21 +01:00

2026-03-03-fxt-overhaul-implementation.md

docs: fxt 2.0 design and implementation plan

2026-03-03 20:11:21 +01:00

2026-03-03-hyperopt-improvements-design.md

docs: add hyperopt improvements design and implementation plan

2026-03-03 18:17:23 +01:00

2026-03-03-hyperopt-improvements.md

docs: add hyperopt improvements design and implementation plan

2026-03-03 18:17:23 +01:00

2026-03-03-ppo-improvements-design.md

feat(ppo): wire symlog, adaptive entropy, percentile scaling into training loop

2026-03-03 20:12:00 +01:00

2026-03-03-rl-agent-gaps-design.md

docs: RL agent gaps design — 7 features for training quality and production readiness

2026-03-03 02:23:38 +01:00

2026-03-03-rl-agent-gaps-implementation.md

docs: RL agent gaps implementation plan — 7 tasks, vertical slices

2026-03-03 02:30:00 +01:00

2026-03-04-argo-migration-design.md

docs: add Argo Workflows migration design

2026-03-04 22:30:12 +01:00

2026-03-04-argo-workflows-phase1.md

feat(argo): add Argo Workflows infrastructure for training orchestration

2026-03-04 22:58:40 +01:00

2026-03-04-argo-workflows-training-design.md

feat(argo): add Argo Workflows infrastructure for training orchestration

2026-03-04 22:58:40 +01:00

2026-03-04-dqn-action-collapse-fix.md

feat(fxt,ml): add fxt train monitor command and update DQN tests for action collapse fix

2026-03-04 22:58:40 +01:00

2026-03-04-dqn-action-collapse-implementation.md

feat(fxt,ml): add fxt train monitor command and update DQN tests for action collapse fix

2026-03-04 22:58:40 +01:00

2026-03-04-gateway-unification-design.md

docs: add gateway unification design (api_gateway + web-gateway → api)

2026-03-04 23:13:35 +01:00

2026-03-04-gateway-unification-implementation.md

docs: add gateway unification implementation plan (22 tasks)

2026-03-04 23:18:10 +01:00

2026-03-04-minio-binary-distribution-design.md

feat(infra): replace PVC binary distribution with MinIO S3 fetch

2026-03-04 16:46:57 +01:00

2026-03-04-minio-binary-distribution.md

feat(infra): replace PVC binary distribution with MinIO S3 fetch

2026-03-04 16:46:57 +01:00

2026-03-04-per-service-ci-design.md

docs: add per-service CI compilation & selective deploy design

2026-03-04 22:58:40 +01:00

2026-03-04-per-service-ci-implementation.md

docs: add per-service CI implementation plan (12 tasks)

2026-03-04 21:56:09 +01:00

2026-03-04-pods-panel-versioning-design.md

feat(infra): standalone Prometheus deployment for foxhunt namespace

2026-03-04 13:24:52 +01:00

2026-03-04-pods-panel-versioning.md

chore: add SubscribeClusterPods stub to trading_service + plan doc

2026-03-04 13:17:41 +01:00

2026-03-04-tui-live-data-design.md

docs: TUI live data wiring design

2026-03-04 01:54:33 +01:00

2026-03-04-tui-live-data.md

docs: TUI live data wiring implementation plan

2026-03-04 02:04:08 +01:00

2026-03-05-hyperopt-degenerate-fix-design.md

chore: track logo, hyperopt design docs, gitignore playwright-mcp

2026-03-05 17:06:53 +01:00

2026-03-05-hyperopt-degenerate-fix.md

chore: track logo, hyperopt design docs, gitignore playwright-mcp

2026-03-05 17:06:53 +01:00

2026-03-06-dqn-action-space-fix.md

fix(ml): DQN hyperopt overhaul — DSR reward, dead features, C51/noisy/network fixes

2026-03-07 12:08:39 +01:00

2026-03-06-dqn-hyperopt-overhaul-design.md

docs: DQN hyperopt overhaul design — 7 remaining root cause fixes

2026-03-06 11:44:02 +01:00

2026-03-06-dqn-hyperopt-overhaul-implementation.md

docs: add DQN hyperopt overhaul implementation plan

2026-03-06 14:32:37 +01:00

2026-03-07-dqn-hyperopt-overhaul-v2.md

fix(ml): DQN hyperopt overhaul — DSR reward, dead features, C51/noisy/network fixes

2026-03-07 12:08:39 +01:00

2026-03-07-gitlab-releases-design.md

feat(infra): GitLab releases with CalVer auto-versioning

2026-03-07 20:38:41 +01:00

2026-03-07-gitlab-releases-plan.md

feat(infra): GitLab releases with CalVer auto-versioning

2026-03-07 20:38:41 +01:00

2026-03-07-hyperopt-objective-fixes.md

fix(ml): 3 hyperopt objective bugs — sliding windows, tanh normalization, adaptive tau

2026-03-07 02:14:31 +01:00

2026-03-07-ml-crate-split-design.md

docs(ml): update plan for 6-crate split (DQN/PPO separate)

2026-03-08 15:13:07 +01:00

2026-03-07-ml-crate-split-plan.md

docs(ml): update plan for 6-crate split (DQN/PPO separate)

2026-03-08 15:13:07 +01:00

2026-03-08-ml-crate-split-phase2-design.md

docs: add ML crate split phase 2 design (domain-aligned extraction)

2026-03-08 19:39:48 +01:00

2026-03-08-ml-crate-split-phase2.md

feat(dqn): enable Branching DQN with 45 factored actions (5×3×3)

2026-03-11 22:00:13 +01:00

2026-03-09-dqn-branching-rainbow-integration.md

feat(dqn): Branching DQN with full GPU Rainbow parity (7 fixes)

2026-03-09 13:17:06 +01:00

2026-03-09-forward-port-main-to-ml-split.md

feat(dqn): enable Branching DQN with 45 factored actions (5×3×3)

2026-03-11 22:00:13 +01:00

2026-03-10-clippy-cleanup.md

fix(clippy): apply cargo clippy --fix across workspace

2026-03-10 11:17:51 +01:00

2026-03-11-cuda-backtest-gpu-residency-plan.md

docs: fix 5 chunk 2 re-review issues in CUDA backtest plan

2026-03-11 09:40:51 +01:00

2026-03-11-cuda-backtest-gpu-residency.md

docs: CUDA backtest & GPU-resident training design spec

2026-03-11 09:11:27 +01:00

2026-03-11-h100-cuda-optimization-hive.md

feat(dqn): enable Branching DQN with 45 factored actions (5×3×3)

2026-03-11 22:00:13 +01:00

2026-05-10-sp21-train-eval-coherence-isv-defrost.md

docs(sp21): T2.2 Phase 1.5 + Phase 2 continuation plan + amendment

2026-05-10 21:35:58 +02:00

2026-05-11-sp21-t22-phase15-phase2-enrichment-realwiring.md

docs(sp21): T2.2 Phase 1.5 + Phase 2 continuation plan + amendment

2026-05-10 21:35:58 +02:00

2026-05-12-sp22-h6-aux-policy-state-bridge.md

docs(sp22): H6 implementation runbook + next-session prompt

2026-05-12 21:07:19 +02:00

2026-05-12-sp22-h6-next-session-prompt.md

docs(sp22): H6 implementation runbook + next-session prompt

2026-05-12 21:07:19 +02:00

2026-05-12-sp22-h6-phase2-recenter-runbook.md

docs(sp22): H6 Phase 2 implementation runbook

2026-05-12 22:27:40 +02:00

2026-05-12-sp22-h6-phase2-recenter.md

docs(sp22): H6 Phase 2 spec — recenter state[121] to [-1, +1]

2026-05-12 22:23:24 +02:00

2026-05-12-sp22-h6-phase3-alpha-beta.md

docs(sp22): H6 Phase 3 α spec revision — atom-position shift (not scalar Q bias)

2026-05-13 01:16:30 +02:00

2026-05-12-sp22-wr-plateau-investigation.md

docs(sp22): H1 falsified + H6 aux→policy state bridge design

2026-05-12 20:58:13 +02:00

2026-05-13-sp22-h6-phase3-alpha-beta-runbook.md

refactor(sp22): H6 Phase 3 α cleanup + runbook revision for atom-shift design

2026-05-13 01:24:15 +02:00