codegen: generated clone-symbol suffixes (__generic/__typed_*/__spec_*/__pshape) collide with user members of the same name #6927

Summary

Perry's specialized/clone symbol naming composes LLVM symbol names by appending a fixed suffix to a user method or function name. A user-declared member whose name already ends in one of those suffixes collides with a generated clone symbol, producing two definitions of the same LLVM symbol.

Repro shape

class C {
  foo(): number { return 1 }
  foo__generic(): number { return 2 }   // collides with the clone of `foo`
}

sanitize_member passes plain [A-Za-z0-9_] names through unchanged, so no escaping prevents the collision.

Scope — the whole suffix family

At least eight generated suffixes share this hazard:

__generic · __typed_f64 · __typed_i32 · __typed_i1 · __typed_string · __typed_f64_recv · __spec_<repmangle> · __pshape (proposed in #6925)

foo + foo__generic already collides on main today, so this is pre-existing and family-wide, not introduced by any single PR.

Severity: low urgency, but should not stay unowned

It fails loudly — duplicate definitions are rejected by the LLVM IR verifier, i.e. a hard compile error, not a miscompile. No program silently takes a wrong path. But the failure mode is inscrutable for a user who simply named a method foo__generic, and the surface grows with every new clone kind.

Suggested fix directions

1. Cheap, local (per-PR mitigation): before emitting a clone, refuse if the composed name is already a registered method/function symbol; stand down to the ordinary guarded lowering (always correct, just slower). perf(codegen): repsel Phase 5a — Ptr<Shape> proven this in methods #6925 plans this for __pshape only.
2. General: make the suffix namespace unforgeable — e.g. reserve a character that sanitize_member escapes out of user identifiers (so a user name can never produce the reserved sequence), or hash-suffix clone symbols the way monomorph/mangle.rs content-addresses specializations, or thread a per-module uniquifier the way func_registry.rs's __dupN disambiguation already does for base symbols.

A general fix touches every clone kind plus the Phase 2 __spec_ reachability ratchet (codegen/spec_abi.rs) and native_proof_regressions, which is why it wants its own change rather than riding along in a perf PR.

Provenance

Found by CodeRabbit review on #6925 (representation-selection Phase 5a); assessed there as real, pre-existing, and family-wide.