Brummel
f7f4c3b237
raw_buf_{int,float,bool}_shadow_rebind_drop_balances_rc_stats guard
the three shipped milestone fixtures (the LLM-natural shadow-rebind
idiom `(let buf (new...) (let buf (set buf...) ...(get buf)))`) with a
live==0 RC-stats assertion alongside their existing stdout check.
This is a SECOND, distinct leak mechanism (A2), separate from the
anonymous-temp path fixed in 62cd4b4 (A1). The A1 call-site drop does
not cover it: here `get`'s arg is the BOUND `buf` (a Term::Var), not an
anonymous temp, so the A1 rule (which excludes Var args) correctly does
not fire, and the leak persists — allocs=2 frees=1 live=1.
Cause (A2). The uniqueness pass runs post-monomorphisation
(infer_module called from codegen lib.rs:993). It registers the
intrinsic monomorph signatures in `globals` under their bare local
name `RawBuf_get__Int` (uniqueness.rs:128), but callee_arg_modes
(uniqueness.rs:390-410) looks them up by the cross-module-qualified
callee-Var name `raw_buf.RawBuf_get__Int` -> MISS -> vec![] -> every arg
defaults to Position::Consume (walk arm uniqueness.rs:240-242). The
borrow-mode RawBuf param of `get`/`size` is thus mis-counted as a
consume, inflating the final `buf` binder's consume_count from 0 to 1,
which suppresses the Term::Let drop gate at codegen lib.rs:1795. The
param_modes in the registered Type::Fn are correct ([Borrow, Implicit]);
the only defect is the name-key mismatch that hides them from the
lookup.
NOT issue #42's "shadow-name collapse": #42's own evidence shows the
distinct-name (b0/b1/...) variant leaks identically, so the defect is
the borrow-counted-as-consume default, not name-keyed ownership.
RED until callee_arg_modes resolves the qualified intrinsic name to the
registered signature. The A1 test (raw_buf_owned_drop_balances_rc_stats)
and the four stdout _e2e tests stay green.