feat: evolve Elevator Queue Lab into decision intelligence workbench

[oosuhada]

Summary

Evolves Elevator Queue Lab from a research simulator/dashboard into a reproducible Decision Intelligence / Operations Research workbench while preserving the simulator and committed evidence as the source of truth.

Architecture

• Keeps ElevatorSimulation, traffic generation, controller logic, event ledger, experiment engine, and learned-policy evaluation intact.
• Adds read-only evidence projections in app/artifacts.py and app/workbench.py for run artifacts, research objects, decision graphs, and deterministic explanations.
• Makes frontend/ the React + TypeScript authored source and keeps web/ as the committed Vite production artifact served by the existing Python server.

Major Features

• Product Shell: Live Operations, Runs, Dispatch Analysis, Experiments, Counterflow Criticality, Models, Explorer.
• Preserved 18-floor / six-car live digital twin, controls, decision inspector, and deterministic replay.
• Versioned PassengerTrace, SimulationRun, DispatchDecision, Experiment, PolicyEvaluation, Model, and TheoryEvidence artifact catalog with provenance.
• Object Explorer and React Flow Decision Trace projected from simulator state, decision history, and event ledger; no graph database.
• Typed semantic ChartSpec validation and registered React renderers.
• Deterministic Ask This Run evidence queries without an LLM source of truth.
• Canonical trace manifest/package support (trace.json, manifest.json, schema.json, validation.json).
• Generalized PR CI, deterministic visual regression, and Playwright showcase capture.

Compatibility

• Preserves python -m app.server --port 4173 as the production runtime.
• Preserves existing /api/health, /api/snapshot, /api/control, /api/replay, /api/experiment, and /api/theory behavior while adding run-scoped workbench APIs.
• Replaces the old authored vanilla browser monolith with the committed React production build without rewriting the simulator core.
• Existing M3/M5/M6/M7 evidence files and deterministic trace identity remain intact.

Evidence Integrity

• No fabricated metrics: analytics use simulator/API state or committed evidence.
• CAPR remains traffic-regime dependent rather than globally superior.
• The fixed M5 RL controller remains a negative/mixed result and is not presented as a general improvement.
• The M7 counterflow threshold remains a fuzzy empirical threshold, not a universal critical constant.
• Synthetic office demand is explicitly not real-building telemetry.

Testing

Local gates after rebasing onto current main:

• TypeScript: PASS
• Vitest: 5/5 PASS
• Vite production build to web/: PASS
• Python unit/contract tests: 51/51 PASS
• Deterministic 2-seed experiment smoke: PASS
• Artifact/trace validation: PASS
• Release audit: PASS
• Playwright functional + responsive + visual E2E: PASS
• Visual suite reproduced successfully across consecutive full E2E runs
• npm run capture:showcase: PASS
• git diff --check origin/main...HEAD: PASS

Visual QA

Committed deterministic visual baselines cover:

• Live Operations desktop
• Live Operations mobile
• Replay
• Experiments
• Theory
• Models
• Explorer / Decision Trace

Committed showcase captures are generated from the real Python backend + production frontend after browser interaction, including a CAPR reassignment and replay flow.

Deployment

The deployment contract remains the existing single Python process serving APIs and web/, compatible with the Mac mini launchd/proxy setup documented under deploy/.

This PR does not claim that the exact M8 commit is already deployed publicly. Production deployment requires the Mac mini to pull the merged commit and pass remote health/browser/release QA.

Known Limitations

• Demand remains synthetic rather than measured building telemetry.
• CAPR is not a global winner across traffic regimes.
• The fixed RL checkpoint does not pass the general-superiority gate.
• The counterflow critical threshold is fuzzy and project-specific.
• Public M8 deployment is pending production infrastructure access/merge; local and CI validation are kept distinct from production verification.

[oosuhada]

Final M8 validation update:

• CI run #76 passed end-to-end on head e27fe00e9e323b817a85afe0b3a18ad90d61b9e1.
• simulator: PASS (51 Python tests, deterministic experiment smoke, destination-control smoke, artifact/trace validation, release audit).
• frontend: PASS (clean install, TypeScript, Vitest, production Vite build).
• browser: PASS (functional E2E, replay, responsive/mobile, Explorer/Decision Trace, deterministic visual-state QA, CI screenshot artifact upload).
• Full statistical regression is intentionally skipped on ordinary PRs and remains gated to main/schedule/explicit label.

Two CI-only issues were fixed without weakening research contracts: the trace-manifest schema was committed so clean checkouts no longer rely on a local-only file, and visual QA now keeps pixel baselines on the controlled local platform while GitHub Ubuntu captures the same seven scenes and enforces required information/geometry/no-horizontal-overflow checks to avoid false failures from OS font rasterization differences.