feat(arigraph): 100%+ Python transcode — prompts, NARS, normalization, spatial

[AdaWorldAPI]

Production prompts ported verbatim from Python AriGraph:

• EXTRACTION_PROMPT: 590 words with examples and edge cases
• REFINING_PROMPT: 400 words with 4 worked examples
• SYSTEM_PLAN_AGENT, SYSTEM_ACTION_AGENT, EXPLORATION_CHECK_PROMPT
• REFLEX_PROMPT for learning from mistakes

Triplet normalization (utils.py:163-172):

• clear_triplet(): I→inventory, P→player, lowercase, strip punctuation
• process_triplets(): digit strip, colon-delimited subject parsing
• parse_triplets_removing(): [[old→new]] format from refinement output

Spatial graph (parent_graph.py:131-203):

• compute_spatial_graph(): auto-extract edges from directional relations
• find_path_directions(): BFS returning movement commands, not locations
• find_direction()/find_opposite_direction(): cardinal direction mapping
• is_spatial_connection(): north/south/east/west detection
• find_unexplored_exits(): identify unvisited directions

Full update cycle (contriever_graph.py:26-84):

• TripletGraph::update(): extract→exclude→refine→delete→spatial
• exclude(): filter known triplets before sending to LLM
• delete_with_location_protection(): won't soft-delete spatial triplets

NARS inference (from adaworldapi/ndarray hpc/nars.rs):

• infer_deductions(): A→B, B→C ⟹ A→C with f=f1f2, c=f1f2c1c2
• detect_contradictions(): same S+O, different relation, both confident
• revise_with_evidence(): repeated observation increases confidence

Fix compilation: collect-before-mutate in compute_spatial_graph,
fix hex literal suffix in nsm_word.rs.

16 new tests for all ported functionality.

https://claude.ai/code/session_01Y69Vnw751w75iVSBRws7o7