fix: prevent simulation destruction on page refresh + add dashboard view

[hassanpasha5630]

Summary

While self-hosting MiroFish-Offline on a 4x GPU Linux server, we discovered that refreshing the browser kills the running simulation and permanently deletes all data. This PR fixes the root cause (two compounding bugs) plus related issues found during debugging, adds a dashboard view to monitor simulations, and fixes orphaned simulation recovery on backend restart.

The kill chain (before this PR)

1. Step3Simulation.vue onMounted() calls doStartSimulation() unconditionally on every page load
2. doStartSimulation() has force: true hardcoded, telling the backend to kill any running process
3. Backend kills the simulation subprocess, deletes all runtime files (action logs, SQLite DBs, run state)
4. Backend starts a new simulation from round 0
5. All previous progress is permanently lost

Bug fixes

• Frontend: check before starting — onMounted now calls getRunStatus() first; if the simulation is already running, it resumes status polling instead of restarting (Step3Simulation.vue)
• Frontend: don't force by default — changed force: true to force: false in the default start path (Step3Simulation.vue)
• Backend: pass GraphStorage to subprocess — the /start endpoint was calling SimulationRunner.start_simulation() without the storage parameter, so GraphMemoryUpdater threw "Must provide storage" and graph updates silently failed. Now fetches neo4j_storage from Flask context and passes it through (simulation.py)
• Backend: preserve state.json on "already running" — when /start returns 400 for an already-running simulation, state.json is now saved with status: "running" before returning the error, preventing a desync where the frontend sees "ready" and retries in a loop (simulation.py)
• Backend: clear Neo4j on force-restart — cleanup_simulation_logs() now accepts optional storage and graph_id parameters to clear stale graph data during force-restart (simulation_runner.py)
• Backend: reconnect to orphaned simulations on restart — when the backend restarts (crash, debug auto-reload, manual restart), the monitor thread that updates run_state.json dies while the simulation subprocess keeps running. On startup, SimulationRunner.reconnect_orphaned_simulations() now scans for simulations with runner_status="running", checks if the PID is still alive, and starts a new monitor thread to resume reading action logs and updating state. Dead processes are marked as stopped. (simulation_runner.py, __init__.py)
• Backend: reconnect GraphMemoryUpdater for orphaned simulations — the orphan reconnect was recovering the monitor thread but not the GraphMemoryUpdater, so graph updates stopped after any backend restart. Now reads graph_id from state.json, creates a fresh Neo4jStorage connection, and restarts the updater so simulation actions continue flowing into the knowledge graph. (simulation_runner.py)
• Backend: fix duplicate graph episodes on reconnect — the orphan monitor started reading action logs from position 0 on every backend restart, re-feeding all actions to GraphMemoryUpdater and creating duplicate episodes in Neo4j. Now starts from end of existing log files so only new actions are processed. (simulation_runner.py)

Additional fixes (found during debugging)

• Bumped default LLM context window from 8192 to 32768 tokens (llm_client.py)
• Added traceback logging for ontology generation failures (graph.py)
• Filter malformed entity/edge types missing the name key before validation (ontology_generator.py)

New feature: Simulation Dashboard (/dashboard)

Built to verify that the bug fixes were working — turned out to be a useful feature:

• Active Now section with live-updating cards for running simulations (progress bars, round counts, stop/view actions, 3s polling)
• All Simulations history table with search input and status filter tabs (All / Running / Completed / Stopped)
• Navigation to Graph, Simulation Run, and Report views per simulation
• Matches existing design language (Space Grotesk, JetBrains Mono, custom CSS, no new dependencies)
• Link added to Home navbar

Test plan

• Start a simulation, refresh the browser — simulation continues running (not killed)
• Dashboard shows running simulation with live progress updates
• GraphMemoryUpdater writes to Neo4j during simulation (verified: 51 entities, 20 relations, 25 episodes)
• Calling /start with force: false on a running simulation returns 400 without killing it
• Restart backend while simulation is running — backend reconnects and resumes monitoring automatically
• Restart backend — GraphMemoryUpdater reconnects and graph updates resume
• Restart backend multiple times — no duplicate graph episodes created
• Dashboard filter tabs and search work correctly
• Navigation from dashboard cards to simulation/report views works