fix: multiple eval improvements and bug fixes

[JackHopkins]

Summary

Inspect AI Sandbox Evaluation

• Add colocated Docker sandbox where Factorio server and FLE Python environment run in a single container managed by Inspect's native sandbox() API
• Bridge architecture: persistent HTTP daemon (bridge_service.py) on Unix socket inside container, thin CLI client (bridge_client.py) invoked via sandbox().exec() from the host-side solver
• Multi-stage Dockerfile: builds FLE wheel, installs on factoriotools/factorio:2.0.73 base with supervisord managing both Factorio and bridge processes
• Sandbox solvers (controlled + unbounded) that mirror the integration solvers but communicate through the bridge
• fle sandbox build CLI command with auto-build on first use of --sandbox
• fle inspect-eval --sandbox --env-id <task> --model <model> to run evals without any external cluster
• --scenario flag to configure which Factorio scenario the sandbox loads

Package Reorganisation

• Move inspect_integration/ -> inspect/integration/
• Add inspect/sandbox/ for sandbox-specific files
• Add shared inspect/eval_set.py with DRY task factory functions (create_throughput_task, create_unbounded_production_task) used by both integration and sandbox eval sets
• Update all imports across the codebase

Agent Namespace Protection

• Freeze namespace names after tool loading so agent code cannot shadow FLE tools (move_to, place_entity, get_entities, etc.) with function or variable definitions
• Raises clear NameError with guidance instead of silently breaking eval

Fix print() Capture Inside Agent-Defined Functions

• Route print() to namespace.log in SerializableFunction.reconstruct() by injecting print=instance.log into the function's globals
• Reorder globals construction: builtins first, then instance attributes, so namespace methods take precedence
• Stop AST print->log rewriting inside FunctionDef bodies (caused infinite recursion when agents defined their own log helper that called print)
• Fix function redefinition bug: read new function from eval_dict (locals) not function_namespace (globals) after exec()
• Increase STDOUT capture limit from 64 to 512 lines; fix truncation to keep first lines instead of last
• Add 12 unit tests covering prints in functions, nested functions, try/except, and the common agent safe() wrapper pattern

Prompt Improvements

• Move policy-writing tips (observe-branch-act, fail fast, diff state, etc.) from task target to system prompt where they belong
• Task target now contains only the concise goal description

Other Improvements

• Add FLE_MODS_DIR / FLE_TOOLS_DIR env var overrides to path resolution in rcon.py
• Handle numpy type serialization in bridge JSON responses (custom _NumpyEncoder)
• Fix Observation.to_dict() round-trip issues (progress="None" string breaking from_dict())
• Graceful fallback when vision rendering fails (no sprites installed)
• Use 2-space indentation instead of tabs in tree observation formatter for consistent rendering
• Auto-start Factorio cluster when no servers are reachable (integration mode)
• Fix ContentImage crash in controlled solver: validate data: URI before passing to ContentImage
• Pass entity dicts instead of repr strings to observation formatter for structured output
• Consolidate redundant sprite-not-found warnings into single message
• Add fastapi to main dependencies and new [inspect] optional group