[guilyx]

What and why

The full code review found a set of bugs where jupyddl returns a wrong answer without complaint: a plan that shouldn't exist, an "optimal" cost that isn't, a heuristic that overestimates or crashes. validate_plan didn't catch any of them, because it replays plans through the same Task.apply that has the bug.

Bug	Before	After
Derived facts never retracted	off deletes (on) but (light) stays true; (not (light)) is true from the start, so bfs "solves" an unsolvable task	Axioms are recomputed from scratch every step and negated derived facts are maintained. Negation inside a rule needs stratification, so it is refused by name
Atom both added and deleted	move(a,a) leaves (at a) and (not (at a)) true, so a 2-step plan "beats" the real 3-step optimum. Same for (and (not p) (when q p))	For complement facts, the delete wins
Actions with no cost effect under a (total-cost) metric	Charged 1, so pegsol and sokoban optimised the wrong objective	Cost 0, as in IPC planners. With no metric, every action still costs 1
(increase (total-cost) 10) inside when	Charged even when the condition is false	Refused
Negative action cost	Accepted, then dijkstra spins until the budget runs out	Refused
Fluent never defined, e.g. (distance a c)	Read as 0, so a road that doesn't exist is free	Conditions and effects that read it fail, as PDDL 2.1 specifies
blind	Skipped zero-cost actions and overestimated, so astar+blind returned cost 3 where 2 is optimal	Takes the min over all costs, including 0
h1/h2/hm	Crash on every numeric or timed task	Fixed
LM-cut with an empty goal	Crash, which also crashed the default solve()	Returns 0
lmcut/h2 admissible flag	Hard-coded True, even with conditional effects	Cleared per task
solve_task(task, "astar")	Used hff (not admissible): cost 6 where 5 is optimal	Optimal planners default to lmcut. -H none meant hff; it now means blind
make_heuristic(LMCutHeuristic, task)	Returned the class unbuilt, so search crashed	Classes and partials are built for the task
Same run, different PYTHONHASHSEED	gripper astar/lmcut expanded 1725, 1758 or 1760 nodes and returned different plans	Fact ids and id sets are built in sorted order, and LM-cut breaks ties explicitly. Benchmarks now reproduce

The requirements registry, the README PDDL table and AGENTS.md now describe the new semantics.

How it was checked

• New tests/test_semantics.py (33 cases): tiny hand-written tasks whose right answer is obvious by inspection. 16 of them fail on the previous code (I ran them against main to confirm) and all pass here. The seed test runs the same search under three PYTHONHASHSEEDs.
• Full suite: 460 passed, 1 skipped. flake8 and black are clean, and web/dist is rebuilt.
• The reviewer's original repro scripts now give the expected answer.

Guarantees changed

• blind is now actually admissible on tasks with free actions.
• lmcut and h2 report admissible=False on tasks with conditional effects; they were already inadmissible there.
• No planner moves in or out of OPTIMAL_PLANNERS in this PR. A follow-up handles bfs/iddfs, which claim cost-optimality but only minimise plan length.

Checklist

• flake8 jupyddl tests tools and pytest pass locally
• python tools/build_web.py re-run; web/dist committed
• Guarantees that changed are listed above
• CHANGELOG.md updated (Unreleased → Fixed)

The history has a merge -s ours of the already-merged #164 tip so the push fast-forwards. The PR diff against main contains only this change. Squash-merge.

---

Generated by Claude Code