Summary

The last two over-acceptances from the §16 review — closing out the correctness wave's planned scope:

• Range non-associativity: the precedence table (docs/reference/operator-precedence.md) marks ../..= assoc "None", but the parser gave them ordinary left-associative binding powers — a..b..c silently parsed as (a..b)..c. Chained ranges now error ("range operators cannot be chained; parenthesize one side") with the span on the second range operator. (a..b)..c stays legal: the parenthesized form is explicit, and the AST cannot (and should not) distinguish it after the fact.
• Item-modifier placement (§16): modifiers were eaten before item dispatch and never validated, so pub async struct S {}, offchain mod m;, pub impl User {}, and pub actor A {} all parsed. Per the grammar, offchain/async prefix only fn_def, and pub is absent from impl_block, actor_def, onchain_mod/event_def, and extern_block. Misplaced modifiers now error on the modifier's own token span, and the item still parses afterward so error recovery stays useful.

Stacked-PR note: targets the integration branch fix/parser-correctness-base (PR #69) — sibling of #70/#72/#73. Shares the usual conflict points with the siblings (tests-module tail, corpus array, AGENTS.md accepted bullet; and #72 adds an identical error_at helper — keep one copy at merge).

Related Issue

None (review-driven; part of the parser correctness wave tracked in PR #69).

Spec Sections Affected

None changed — implements the precedence table's "None" associativity row and §16's modifier placement as written.

Checklist

• Code follows the Sploosh design principles (one way to do it, explicit over implicit, etc.)
• Documentation updated in relevant docs/ pages (crates/AGENTS.md subset contract)
• Tests added or updated
• All build targets still compile (if applicable)
• Spec-only PR (skip Build Targets section if checked)

Build Targets Tested

• N/A (docs/spec only) — no codegen targets exist yet; "build" = the Rust workspace.

Test Plan

• 4 new unit tests: range_operators_cannot_chain (a..b..c, a..=b..c, a..b..=c all rejected with the chaining message), single_and_parenthesized_ranges_parse (lo..hi, lo..=hi, (a..b)..c accepted), offchain_and_async_apply_only_to_fn_items (five rejected placements + pub offchain async fn accepted), pub_rejected_where_grammar_omits_it (rejected on impl/actor/onchain/extern; accepted on struct/enum/trait/mod/use/const/type).
• New corpus fixture tests/corpus/ranges_modifiers.sp covering the accepted side (ranges in expressions, pub declarations, pub offchain async fn).
• cargo fmt --all -- --check, cargo clippy --workspace --all-targets -- -D warnings, cargo test --workspace (21 tests) green locally and via scripts/docker-check.ps1 -Build (Ubuntu parity).

Summary by cubic

Enforces non-associativity for ../..= and validates item modifier placement per §16 to close parser over-acceptances. Chained ranges now error, and misplaced pub, offchain, and async are rejected with clear, anchored diagnostics; tests assert the error spans.

• Bug Fixes

  • Reject chained ranges like a..b..c; allow (a..b)..c.
  • Restrict offchain/async to fn items.
  • Reject pub on impl, actor, onchain, and extern.
  • Diagnostics point at the second range operator or the misplaced modifier; item still parses for recovery.

• Migration

  • Parenthesize if chaining was intended: (a..b)..c.
  • Use offchain/async only on fn, and remove pub from impl/actor/onchain/extern.

^{Written for commit 36c17f1. Summary will update on new commits.}