parser: reject builtin-named binding LHS with ILO-P011

[danieljohnmorris]

Summary

Pattern-matches PR #245's parse-time rejection of builtin-named user functions, extending it to local-binding LHS. A binding like flat=cat ls " " was silently accepted by the parser; any later use of flat in operand position resolved to the 1-arg flatten builtin (the verifier checks is_builtin before locals), and the agent saw the misleading ILO-T006 'flat' expects 1 args, got 0 instead of a signal that the local binding was being shadowed.

Persona origin: 2026-05-16 pdf-analyst re-run against v0.11.2, friction #6.

Manifesto framing: silent mis-dispatch is the worst class of bug for an AI agent because there is no diagnostic to retry against. The parser now intercepts at the binding site with ILO-P011 \flat` is a builtin and cannot be used as a binding name plus a rename hint (myflat, flatv`). First retry is the right one.

Repro

Before:

$ ilo 'main>n;flat=5;flat' --run-tree main
ILO-T006 arity mismatch: 'flat' expects 1 args, got 0

After:

$ ilo 'main>n;flat=5;flat' --run-tree main
ILO-P011 `flat` is a builtin and cannot be used as a binding name
hint: rename to something like `myflat` or `flatv`. Builtins shadow local bindings in call position, so reusing the name silently mis-dispatches.

What's in the diff

• f98748e parser: reject builtin-named binding LHS with ILO-P011 — adds Builtin::is_builtin(name) guards to both parse_decl (top-level) and parse_stmt (in-function) binding sites, runs after the existing per-name fld= block so the fold-specific hint still wins. Twelve-builtin cross-engine regression test (tree, VM, Cranelift) covering both top-level and in-function binding forms, the exact pdf-analyst rerun3 repro, and a sanity check that the renamed binding compiles and runs. examples/builtin-binding-name-rename.ilo demonstrates the correct shape.

• 3547b6b docs: note builtin names rejected as binding LHS — extends the Reserved words section in SPEC.md (auto-regenerates ai.txt via build.rs) and the syntax-rules section in skills/ilo/SKILL.md to mention the new rejection, listing the same names the parser checks and both the function-name and binding-LHS error shapes.

Test plan

• cargo test --release --features cranelift passes the full suite (127 test crates, 0 failures) in the isolated worktree target dir
• New regression_builtin_binding_name (15 tests) passes on tree, VM, and Cranelift
• Existing regression_friendly_errors::friendly_fld_binding still passes (per-name fld block preserved)
• Existing regression_builtin_fn_name (PR reject builtin-named user functions at parse time #245 parent) still passes
• examples_engines.rs exercises the new builtin-binding-name-rename.ilo across every engine and confirms len parts == 3
• git diff ai.txt matches the SPEC.md-driven regeneration (build.rs CI check passes)

Follow-ups

• Destructure pattern names ({flat;y}=...) and @flat xs{...} foreach variables are out of scope. They share the same operand-position dispatch risk but are much rarer in practice and would each need a separate intercept point. Filing as parked.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ All tests successful. No failed tests found.

📢 Thoughts on this report? Let us know!