[ghostfix-pm]

Summary

Four defects in the diagnostic paths, all of the same shape: pyval knew something was wrong and told the caller nothing. Each one is reproduced on the parent commit below, and the fixes come with regression tests. No packaging, no CI, no dependency changes.

Bug fixes

1. An unevaluable step crashed validation with a traceback. Numeric fluents have no default in PDDL, and the simulator raises UPUsageError when a condition or effect reads an instance :init never assigned. simulate let that escape, so the CLI died mid-report:

$ pyval init_domain.pddl init_problem.pddl init_plan.txt
unified_planning.exceptions.UPUsageError: The state {at(t1, l2): true, fuel(t1): 40, ...}
does not have a value for the value fuel(t2)

The same input now stops at the offending step, names the fluent and the repair, and refuses to give a verdict it did not earn:

Step 2: (drive t2 l1 l2) ✗ EVALUATION ERROR
  No value in the current state for fuel(t2): this problem's :init never assigns it,
  and PDDL defines no default. The step cannot be evaluated until it is initialised.
  Add (= (fuel t2) <value>) to :init.

=== Goal Check ===
Plan could not be validated: step 2 of 2 ((drive t2 l1 l2)) could not be evaluated,
so no verdict is given for the plan as a whole.

2. ✗ PRECONDITION FAILURE with nothing under it. decompose_preconditions decided or by counting collected sub-failures against the number of disjuncts, so (or (and (p) (q) (s)) (r)) produced four failures for two disjuncts and reported none; and any node it did not handle (forall, implication, object equality) fell straight through. Both produced a failed step with an empty unsatisfied_preconditions list. Disjuncts are now evaluated with StateEvaluator and an unhandled node gets a final branch, so:

Step 1: (go a) ✗ PRECONDITION FAILURE
  Unsatisfied: ((p(a) and q(a) and s(a)) or r(a))
    None of the 2 disjuncts holds: (p(a) and q(a) and s(a)); r(a)

Object equality (:equality) also no longer goes through the numeric path, where it answered "Could not evaluate numeric comparison" with type: "numeric".

3. Metrics silently disappeared. _evaluate_metric used state.get_value, which asserts on arithmetic sub-expressions — the same root cause 0.1.4 fixed for goals, left behind on the metric path. (:metric minimize (+ (* 2 (fuel l1)) 5)) returned no metric at all (expected 17.0), and :action-costs domains were dropped entirely because MinimizeActionCosts is not an expression over the final state. Both are reported now, and a metric that genuinely cannot be evaluated shows as n/a with the reason.

4. --track never matched anything. The README's own command pyval domain.pddl problem.pddl plan.txt --trajectory --track fuel printed "No numeric fluents to display.": the tracker compared a display name (fuel truck1) with the snapshot key (fuel(truck1)), and format_trajectory compared the raw reference with that same key. Both now accept the bare function name (every instance), the PDDL spelling and UPF's rendering; an unmatched reference warns with the fluents that do exist instead of printing an empty table.

Also fixed

The Phase-1 "numeric function has no initial value" warning was decided per declared name, so initialising (fuel t1) silenced it for fuel(t2), and it claimed the missing value "defaults to 0" — which is not what the simulator does (it raises). It now enumerates ground instances and reports the ones an action writes, with a valid PDDL repair hint. Static tables such as (distance ?a ?b) are left alone when they are partially specified, since that is normal.

Additive output changes

• JSON gains is_valid, phases.goal_check (every goal and its verdict) and warnings; a failed step may carry evaluation_error. phases.goals is unchanged, so existing consumers keep working — goal_check is what distinguishes "checked and green" from "not checked", which goals: null could not.
• ValidationResult gains goal_results (all goals) alongside the existing unsatisfied_goals.

Tests

107 passed, 8 skipped in 17.2s

40 new tests (67 → 107), each defect pinned by a test that fails on the parent commit. Text, verbose and trajectory output for every existing fixture is byte-identical to the parent commit; the JSON diff is purely additive (verified by diffing full reports for six fixtures before and after).

Files Changed

File	Δ	What
pyval/plan_simulator.py	+152/-30	Catch evaluation failures into StepResult.evaluation_error; StateEvaluator metrics, :action-costs sum, unevaluable metric reason
pyval/diagnostics.py	+177/-52	Disjuncts evaluated instead of counted; final branch for quantifiers/equality; :equality off the numeric path
pyval/syntax_checker.py	+104/-11	Per-instance numeric :init coverage, capped enumeration, PDDL repair hints
pyval/report_formatter.py	+68/-5	EVALUATION ERROR step lines, "could not be validated", n/a metrics, is_valid/goal_check/warnings, unified --track filter
pyval/numeric_tracker.py	+76/-7	fluent_reference_matches/parse_fluent_reference, unknown-reference tracking
pyval/validator.py	+38	Warning for --track values that select nothing
pyval/models.py	+8	StepResult.evaluation_error, ValidationResult.goal_results
pyval/pddl_format.py	+24	as_pddl_call: fuel(t2) → (fuel t2) for repair hints
tests/test_numeric_init.py	+226	8 tests
tests/test_precondition_diagnostics.py	+192	7 tests
tests/test_evaluation_errors.py	+147	6 tests
tests/test_metrics.py	+131	5 tests
tests/test_tracking.py	+119	9 tests
tests/test_json_output.py	+71	4 tests
README.md	+12	Numeric initialisation, evaluation errors, accepted --track spellings
CHANGELOG.md	+88	[Unreleased] entries