feat: add Atomizer HQ multi-agent cluster infrastructure

- 8-agent OpenClaw cluster (Manager, Tech-Lead, Secretary, Auditor,
  Optimizer, Study-Builder, NX-Expert, Webster)
- Orchestration engine: orchestrate.py (sync delegation + handoffs)
- Workflow engine: YAML-defined multi-step pipelines
- Agent workspaces: SOUL.md, AGENTS.md, MEMORY.md per agent
- Shared skills: delegate, orchestrate, atomizer-protocols
- Capability registry (AGENTS_REGISTRY.json)
- Cluster management: cluster.sh, systemd template
- All secrets replaced with env var references

hq/workspaces/manager/memory/2026-02-10.md

@@ -0,0 +1,135 @@
+# 2026-02-10
+
+## Hydrotech Beam — KBS Sessions Received
+- Antoine recorded 3 KBS capture sessions on his Windows machine (NX/Simcenter)
+- Data location: `/home/papa/ATODrive/Projects/hydrotech-beam/Hydrotech-Beam/_capture/`
+- Sessions: `20260210-132817` (6s), `20260210-161401` (38s), `20260210-163801` (414s main session)
+- Main session is a full walkthrough of the NX model with parameter names, values, BCs, materials
+
+### New Information from KBS Sessions
+- Beam length = 5,000 mm (`beam_length` expression)
+- Cantilever: left fixed, right loaded with 10,000 kgf downward
+- Hole span = 4,000 mm (`p6`), holes start/end 500mm from beam ends
+- Mass via expression `p1` (NOT `p173` as we had) — starting 11.33 kg (CONTRADICTS 974 kg baseline!)
+- Material: ANSI Steel 1005 — future: aluminum 6061, stainless ANSI 310
+- Mesh: CQUAD4 thin shells, mid-surface idealization, element size = 67.4/2
+- New expression names: `beam_half_height`, `beam_half_width`
+- `p6` (hole span) as potential new design variable
+- 4 screenshot triggers in the session metadata
+
+### Actions Taken
+- Posted acknowledgment + next steps in #project-hydrotech-beam
+- Spawned Tech Lead sub-agent (label: tech-lead-kb-update) to:
+  - Process all 3 transcripts
+  - Update KB to Gen 002
+  - Reconcile mass discrepancy (11.33 kg vs 974 kg)
+  - Close resolved gaps (G1, G2, G5 partial, G8)
+  - Update CONTEXT.md
+  - Commit to Gitea
+
+### Workflow Status
+- Step 1 (Optimizer strategy): OPTIMIZATION_STRATEGY.md exists as DRAFT from Feb 9
+- Current: Processing new KB data before proceeding
+- Next: Optimizer revises strategy with confirmed params → Auditor review → Study Builder code
+- Model files confirmed synced: Beam.prt, Beam_fem1.fem, Beam_fem1_i.prt, Beam_sim1.sim
+
+### Completed
+- [x] Tech Lead completed KB Gen 002 update — commit `b88657b`
+- [x] Mass corrected AGAIN: **1,133.01 kg** (`p173`), NOT 11.33 kg — Antoine corrected us
+- [x] Binary introspection of Beam.prt — extracted complete expression table (commit `15a457d`)
+- [x] DV baselines are NOT round: face_thickness=21.504, core_thickness=25.162 (not 20/20)
+- [x] Gaps G12-G14 closed (beam_half_height=250, beam_half_width=150, holes_diameter expression confirmed)
+- [x] Important: `beam_lenght` has TYPO in NX (no 'h') — scripts must use exact spelling
+- [x] `hole_count` links to `Pattern_p7` in the NX pattern feature
+- [x] CONTEXT.md updated with full expression map, pushed to Gitea
+
+### Pending — Waiting on Antoine
+- [ ] Baseline re-run (G10, G11) — need current displacement and stress values
+- [x] Decision on `p6` (hole span) — kept fixed at 4,000mm for now (Manager decision)
+
+### Windows Environment (dalidou)
+- Path: `C:\Users\antoi\Atomizer\projects\hydrotech-beam\` (Syncthing from server)
+- Python: `anaconda3\envs\atomizer` (conda env named "atomizer")
+- Antoine ran smoke test on Feb 11 — hit 2 bugs, both fixed (commit `135698d`)
+- NXOpen implementation still needed (solve, extract_displacement, extract_stress)
+
+### In Progress
+- [x] Optimization strategy updated with corrected baselines (commit `3e51804`)
+- [x] Auditor review: APPROVED WITH CONDITIONS — 2 blockers found and fixed:
+  - Hole spacing formula: `span/(n-1)` not `span/(n+1)` — fixed
+  - Web height constraint: added `500 - 2*face - dia > 0` pre-check — fixed
+  - Commit `94bff37`
+- [x] Study Builder completed Phase 1 code (commit `017b90f`) — verified end-to-end with stub solver
+  - 6 files: run_doe.py, sampling.py, geometric_checks.py, nx_interface.py, requirements.txt, README.md
+  - Pre-flight geometric filter catches ~24% of infeasible combos
+  - NXOpen template ready — needs 3 methods filled in on Windows (solve, extract_disp, extract_stress)
+- [ ] Antoine running baseline SOL 101 for displacement + stress (parallel)
+- [ ] `p6` kept fixed at 4,000mm for now (DEC by Manager)
+
+### NXOpenSolver → Existing Engine Integration (Late Evening)
+- Antoine confirmed: runs everything from his "Honda atomizer" conda env on Windows
+- Uses existing `run_optimization.py` which calls `NXSolver` + `NXParameterUpdater` + pyNastran extractors
+- **Key insight:** We do NOT need to write NXOpen code from scratch — the Atomizer engine already has everything:
+  - `optimization_engine/nx/solver.py` — journal-based solver via `run_journal.exe`
+  - `optimization_engine/nx/updater.py` — expression updates via `.exp` import
+  - `optimization_engine/extractors/extract_displacement.py` — pyNastran OP2
+  - `optimization_engine/extractors/extract_von_mises_stress.py` — pyNastran OP2, kPa→MPa
+  - `optimization_engine/extractors/extract_mass_from_expression.py` — from temp file
+- Delegated to Study Builder (label: `study-builder-nx-impl`) to rewrite `NXOpenSolver` as a wrapper around existing engine
+- Asked Antoine to confirm `pyNastran` is installed in the conda env
+
+### Infra Fixes
+- Study Builder model was set to non-existent `claude-sonnet-5` → fixed to `claude-sonnet-4-20250514`
+- All agents were missing Anthropic API auth → propagated from Manager's auth-profiles.json
+- Agents fixed: secretary, study-builder, optimizer, auditor, technical-lead
+
+### Study Builder Delivered — NXOpenSolver (commit `33180d6`)
+- Wraps existing Atomizer engine: NXParameterUpdater, NXSolver, pyNastran extractors
+- HEEDS-style iteration folders, 600s timeout, CQUAD4 shell stress, kPa→MPa
+- Full interface compatibility with run_doe.py preserved
+- 252 additions, 126 deletions
+
+### Tech Lead Refined — NXOpenSolver v2 (commit `390ffed`)
+- Built on Study Builder's work with improvements:
+  - Element type auto-detection (tries solids first, falls back to CQUAD4)
+  - OP2 fallback path (solver result → expected naming convention)
+  - Mass fallback via `_temp_part_properties.json`
+  - Follows SAT3_Trajectory_V7 FEARunner pattern exactly
+- Both commits stack cleanly on main, latest is the active version
+
+### Late Night — Antoine Follow-Up (~23:00-01:00 UTC)
+- Antoine returned: "Yeah! What's next?" — confirmed ready to move forward
+- Asked about conda env: confirmed he uses `conda atomizer` (defined in `environment.yml` at repo root)
+  - Includes optuna, scipy, numpy, pandas, pyNastran — all Phase 1 deps covered
+- Asked "What's the NXOpen implementation about?" — explained the 3 bridge methods (solve, extract_disp, extract_stress)
+- Antoine asked how this relates to legacy Atomizer studies (SAT3, mirror blank)
+  - Confirmed: same engine (NXSolver, NXParameterUpdater, pyNastran extractors)
+  - Differences: geometric pre-filter, LHS sampling, cleaner separation, project-scoped
+- **Antoine approved:** "go ahead and do it"
+- Delegated NXOpen implementation completion to Technical Lead (label: `hydrotech-nxopen-impl`)
+  - Task: complete NXOpenSolver.evaluate() using existing Atomizer engine components
+  - Reference: SAT3_Trajectory_V7, bracket study, existing engine classes
+
+### Feb 11 Morning — Bug Fixes + Final Refactor
+- Antoine tested on dalidou, hit 2 bugs:
+  1. SQLite duplicate study name → fixed with `load_if_exists=True` + `--clean` flag
+  2. Sampling crash with `n-samples 1` → skip stratified patching when n < 11
+  - Commit `135698d`
+- **Full refactor of nx_interface.py** (commit `126f0bb`):
+  - `AtomizerNXSolver` wraps existing `optimization_engine` (NXSolver + pyNastran extractors)
+  - HEEDS-style iteration folders, .exp file generation, OP2 extraction
+  - StubSolver improved with beam-theory approximations
+  - Windows path confirmed: `C:\Users\antoi\Atomizer\projects\hydrotech-beam\`
+  - Conda env: `atomizer` (all deps pre-installed)
+
+### Future Initiative — NX Simcenter 3D MCP (CEO request, Feb 11)
+- MCP server on dalidou for direct NXOpen interaction
+- Endpoints: expressions.list/get/set, model.update, solve.run, results.*, introspect, screenshots
+- Eliminates pyNastran, temp files, journal generation — all via NXOpen API
+- Target: Phase 2/3 roadmap
+- Logged per Antoine's explicit request — not blocking current work
+
+### Next
+- [ ] Antoine tests `--backend nxopen` on dalidou (single trial smoke test)
+- [ ] Full 51-trial Phase 1 run
+- [ ] Phase 2 TPE optimization