docs(spec): bounded-exhaustive verifier for parser vs grammar drift (#643) - #644

Design spec for #643: a deterministic, bounded-exhaustive verifier that checks the real parser against tests/grammar/grammar.pest, plus generating the EBNF from grammar.pest.

Spec: docs/superpowers/specs/2026-10-20-grammar-drift-verifier-design.md

Settled with the maintainer (2026-10-20):

• syntax/semantic classification of error codes is a test-only list (no public API);
• the EBNF is a generated docs/GRAMMAR.md, published by the docs site as a snapshot page;
• main gets the verifier only, with no new error codes.

Summary:

• §3: 24 user-facing INT-019 sites and the INT-000 retract-arity path get their own PRS codes; 4 unreachable INT-019 sites stay.
• §4: property: pest rejects ⇒ parser rejects; pest accepts ⇒ parser accepts or rejects with a semantic code. Semantic codes need a witness and one meaning per code.
• §5: lexical level (all ≤3-char strings, ~1.07M) and structural level (12 templates × balanced token sequences, k = 3 default, deeper on a schedule); KNOWN_DRIFT predicates, each failure matching exactly one.
• §7: pest → EBNF renderer with a staleness test.
• §9: PR order.

Reviewed with Codex; its findings (witness rule, overlapping drift predicates, token count, balance definition and cost, ordering) are addressed in the second commit.

Docs only; no code change. Part of #643.

🤖 Generated with Claude Code

https://claude.ai/code/session_019kY61Y11gcRD91Y6b7jbGT