Gsm v2

[emuskardin]

Blocking

1. has_local_compatibility / has_score_function are inverted — aalpy/learning_algs/general_passive/ScoreFunctionsGSM.py:91
Both return True when the method is not overridden. Combined with the guard at line 170, SimpleFutureBasedCompatibility(local_compatibility=..., ...) always raises.
Repro: run_GSM(data, transition_behavior="stochastic") → ValueError("Exernal local compatibility is provided…") at GeneralizedStateMerging.py:130. Also breaks Examples.example_Alergia_extension and gsm_IOAlergia_domain_knowldege.

2. CountDataHandler.copy drops the defaultdict — aalpy/learning_algs/general_passive/DataHandler.py:264
copy rebuilds transition_count as a plain dict, but merge (line 257) relies on defaultdict (x.transition_count[in_sym]). Any reversible (first-pass) partitioning over count data crashes.
Repro: run_GSM(..., transition_behavior='stochastic', score_calc=SimpleScoreCalculation(score_function=...)) → KeyError: 'b'.

3. add_labeled_sequence never calls init_data() — aalpy/learning_algs/general_passive/DataHandler.py:104
Children are created as GsmNode((in_sym, out_sym), curr_node) without self.init_data(), unlike add_trace at line 76, so data is None.
Repro: CountDataHandler().createPTA(labeled_sequences) → AttributeError: 'NoneType' object has no attribute 'transition_count'.

4. Deterministic compatibility is no longer enforced — aalpy/learning_algs/general_passive/GeneralizedStateMerging.py:385
The removed compute_local_compatibility enforced GsmNode.deterministic_compatible for transition_behavior="deterministic" regardless of score_calc; the replacement only checks Moore compatibility, and only when output_behavior == "moore". The same hunk downgrades the nondeterministic-input check to warnings.warn (line 179), removing the second guard.
Repro: a mealy/deterministic run with a user-supplied score_calc (Examples.gsm_edsm, or any SimpleScoreCalculation(score_function=...)) yields a root with 'a': ['x','y'] and is_deterministic() == False, which to_automaton then silently collapses.

5. CacheSUL.__getattr__ recurses infinitely — aalpy/base/SUL.py:136
Recurses whenever self.sul is absent, e.g. the __setstate__ lookup during unpickling.
Repro: pickle.loads(pickle.dumps(CacheSUL(AutomatonSUL(dfa)))) → RecursionError.
Fix: guard with if name == 'sul': raise AttributeError(name) (or use object.__getattribute__).

Correctness

6. best_candidate update is nested in the cache-miss branch — GeneralizedStateMerging.py:216
The best_candidate / best_score update sits inside if partitioning is None:. Cached partitionings survive promotions (only merges clear the cache), so from the round after a promotion onward a cached candidate can never be selected as best — even one with ImmediateAccept. The best is then chosen only among freshly computed pairs, giving a different/worse merge sequence than intended.

7. ScoreCombinator fakes an early verdict — ScoreFunctionsGSM.py:329
initialize_merge aggregates sub-results with aggregate_score, whose default returns a list. Sub-scores with no early verdict return None, so the result is [None, None], which is not SpecialScores.NoScore. GSM treats that as an early verdict, never builds the partitioning, and never calls the combinator's local_compatibility or score_function — every ScoreCombinator accepts merges on a meaningless score.

8. CountOnPTADataHandler.copy loses PTA state — DataHandler.py:275
It inherits copy from CountDataHandler, which returns a bare CountData, dropping pta_count and shadow_pta and changing the type. When such a copy is written back (GeneralizedStateMerging.py:250), later hoeffding_compatibility(compare_original=True) raises AttributeError: pta_count, and SimpleFutureBasedCompatibility raises "no PTA data is available". Reachable whenever CountOnPTADataHandler is combined with a score calc that has no early verdict.

9. EDSM_score checks are dead — GsmAlgorithms.py:40
It still tests node.get_prefix_output() is None to skip "no evidence" nodes, but unknown_output changed from None to object() (GsmNode.py:32). Both checks (lines 40 and 44) are now dead, so unlabeled nodes count as evidence and run_EDSM's scoring differs from its intent.

10. Stale cached partitioning misses newly-blue successors — GeneralizedStateMerging.py:241
After a promotion only the promoted blue's cache entries are dropped, but red_states_backing_set has grown. Cached partitionings captured red_partitions at first-pass time; when such a candidate is later applied the second pass returns early and reuses the stale new_blue list, so successors that should have become blue are never enqueued. Currently masked by finding 6 — it goes live once that is fixed.

11. moore_check uses red instead of partition — GeneralizedStateMerging.py:386
It calls GsmNode.moore_compatible(red, blue) on the original red rather than on partition, whose prefix output may already have been resolved by resolve_unknown_prefix_output. The check is strictly more permissive than intended and can let a Moore-incompatible implied merge through.

API / reporting

12. run_Alergia_EDSM dropped from package exports — aalpy/learning_algs/__init__.py:15
Replaced by run_Alergia_GSM, so from aalpy.learning_algs import run_Alergia_EDSM — previously public API — now fails. The alias survives only inside GsmAlgorithms (line 160), which is what the test imports.

13. nr_merged_states is double-counted — GeneralizedStateMerging.py:381
Incremented in both passes, so pairs deferred to remaining_merges are counted twice. ProgressReport.log_merge (Instrumentation.py:98) accumulates this, making the printed "Remaining" count drift and go negative.

14. First broken promotion goes unreported — Instrumentation.py:154
In MergeViolationDebugger.log_promote the new elif node is None branch is unreachable for the first broken promotion: self.map.get(None) is None, so the old_red is None branch fires first, inserts None as a map key and logs a normal "promote". Only the second broken promotion is reported.