Rewrite jupyddl as a pure-Python PDDL planning framework (remove Julia)

[guilyx]

Overview

Removes the Julia / PDDL.jl / PyCall / pyjulia dependency entirely and rewrites jupyddl from scratch as a pure-Python, zero-runtime-dependency planning framework: a hand-written PDDL parser + grounder, a suite of planners (classical → SOTA), a family of heuristics, a benchmarking harness, and a CLI — with a comprehensive test suite.

What's included

Parser (jupyddl/parser/) — tokenizer, AST and recursive-descent parser for STRIPS, :typing (with hierarchy), :negative-preconditions, :equality, :action-costs ((increase (total-cost) k)), universally-quantified goals, and forall/when conditional effects. Numeric fluents beyond total-cost are rejected with a clear UnsupportedFeatureError.

Grounding (jupyddl/grounding.py, task.py) — type-aware instantiation, static-predicate pruning, positive-normal-form compilation of negative preconditions/goals, object harvesting for undeclared constants, and conditional/quantified effect expansion into an integer-encoded grounded Task.

Planners (jupyddl/search/) — bfs, dfs, iddfs, dijkstra, gbfs, astar, wastar, idastar, ehc (enforced hill climbing), on a shared best-first engine, behind a registry.

Heuristics (jupyddl/heuristics/) — blind, goalcount, hmax, hadd, hff (FF relaxed plan), critical-path h^m (h1/h2), and LM-cut, behind a registry.

Framework glue — high-level API (solve, build_task, solve_task, validate_plan), a benchmarking harness (jupyddl/benchmark.py, CSV + plots) and a CLI (jupyddl solve / jupyddl benchmark).

Packaging — migrated to pyproject.toml (hatchling, uv); core has no runtime dependencies (matplotlib is an optional viz extra). CI reworked to a pure-Python matrix (3.9–3.12), no Julia.

Correctness highlights

• Optimal planners (bfs, dijkstra, astar/idastar with an admissible heuristic) agree on optimal cost on every solvable example; every planner returns a validated plan.
• Admissible heuristics (blind, hmax, h1, h2, lmcut) never overestimate on STRIPS instances; h1 == hmax; lmcut >= hmax.
• dinner (op2(a,a,c)) and flip (3 flips) reproduce the previous Julia implementation's plans.

Known scope notes

• On conditional-effect domains (flip), the delete-relaxation heuristics (hadd/hff/lmcut/h^m) are satisficing only (not guaranteed admissible), since each conditional effect is relaxed into its own operator. Use bfs/dijkstra or astar/idastar with blind for guaranteed-optimal plans there. This matches how tools like Fast Downward restrict LM-cut to STRIPS.
• Example-data quirks (external submodule, unchanged): grid is numeric (unsupported); vehicle has typos (struck/truck, acessible) making its goal unreachable. Both are handled gracefully.

Testing

• ✅ pytest --cov=jupyddl — 92 passed, ~92% coverage
• ✅ flake8 jupyddl tests — clean (config in .flake8)
• ✅ jupyddl solve and jupyddl benchmark CLIs

Demos

CLI solve (optimal A*+LM-cut on tsp, and A*+blind on the conditional-effect flip):
cli_solve_demo.txt

Comparative benchmark across planners (node expansions; informed search expands far fewer nodes):
Planner comparison (expanded nodes)

Full per-run benchmark CSV:
benchmark_results.csv

Notes

AGENTS.md documents the new pure-Python environment and the non-obvious scope notes above.

_{To show artifacts inline, enable in settings.}

  

[mergify]

This pull request is now in conflicts. Could you fix it @guilyx? 🙏