Bound ASPIC property-test theories and pin the #102 argument budget

[ctoth]

This PR contains test-only safeguards for #102 and makes no production change. The production fix, rule canonicalisation (option A in the investigation comment), is waiting on a semantic decision, so #102 stays open.

Problem

The ASPIC rule generators in tests/structured/aspic/test_aspic.py drew rule antecedents with replacement. One seed such as ~q, ~q -> p, together with its positional transpositions (~p, ~q) -> q and (~q, ~p) -> q, multiplied a 4-premise theory into 18,850 arguments and roughly 306M attack pairs. Once Hypothesis saved that example in a local database, it replayed on every run: the test hit the 600s timeout, and pytest then raised a MemoryError while formatting the traceback.

Changes

• strict_rules, strict_seed_rules and defeasible_rules now draw distinct antecedents in canonical (sorted) order, using _distinct_antecedents.

• knowledge_base and well_defined_knowledge_base reject theories with more than 2,000 arguments, via assume(...). The count is taken from build_arguments with an empty contrariness function. Because the c-consistency filter only ever removes arguments, that count is an upper bound on the real argument set. As a result, any saved example, including an old one, is discarded rather than run through compute_attacks.

• tests/structured/aspic/test_aspic_argument_budget.py pins the ASPIC compute_attacks blows up on cyclic strict rules: 4 premises, 14 rules give 18,850 arguments and ~8.7M rebuttals #102 theory with a deterministic operational contract (AGENTS.md):

  • at most 200 arguments;
  • at most 20,000 attack pairs, counted from per-conclusion argument counts, without calling compute_attacks.

  It is marked xfail(strict=True) until option A lands. It fails in about 0.5s instead of timing out. A control confirms that a small theory with distinct antecedents stays within budget.

Refs #102

🤖 Generated with Claude Code

https://claude.ai/code/session_01V4tAVyyKcs1sYcsEzL7Bzg