gen-c: 5912 errors are symbols referenced and never declared, and cast_i8 alone is 210

[gHashTag]

What the measurement says

I set out to type lists of calls, estimating ~92 errors from the shape count. The change removed one. The reason is the finding:

$ grep -rl 'cast_i8(' specs/ | wc -l      # 6 specs call it
$ grep -rl 'fn cast_i8'  specs/ | wc -l   # 0 declare it

fn_return_types has no entry, so a lookup-based inference correctly refuses. That refusal is not the problem — the missing declaration is.

Counted over the generated C corpus with -ferror-limit=0:

class	count
use of undeclared identifier	3007
call to undeclared function	2057 (91 files)
unknown type name	848
total	5912

against 522 remaining in the __auto_type class I have spent three passes on. The undeclared-symbol family is more than ten times larger, and it was never counted because I was following one class rather than looking at the distribution.

Most-called undeclared functions:

210  cast_i8      142  len          76  compose      64  booth_mul_i32
 62  approximately_equal            58  adder_tree   52  random_input
 50  ternary_gemm

cast_i8 is not a user function

The compiler knows it. compiler.rs:9732 carries the note "integer cast. cast_i8( alone appears 1,100 times", and compiler.rs:5585 records that it is "never lowered ('use of undeclared identifier 'cast_i8'' from inside an emitted .{ cast_i8(42) })".

So it is a convention the parser understands and at least one backend does not emit. len is the same shape — Rust's .len() has a lowering, a bare len does not.

What has NOT been measured

Whether these split into causes: a compiler-known builtin never emitted (cast_i8, len), a function declared in another spec and reached through use, or a function declared nowhere at all. Those need different repairs and the counts above do not separate them. That separation is the next measurement, and it should come before any code — the same first move that turned 1729 __auto_type errors into three tractable problems and one impossible one.

What this issue does not claim

That 5912 errors would disappear. An undeclared symbol often cascades: one missing declaration silences and then reveals diagnostics about every use. The number is a population, not a prediction.

[gHashTag]

Split by cause, which the totals could not do. tools/c_error_classes.py --undeclared (added in #3466):

	count	cause	examples
A	487 (23%)	declared as fn in some spec, not reaching this file	len(142), mean(40), decode(26), all(24)
B	149 (7%)	the name of a module or spec file, called as a function	adder_tree(58), ternary_gemm(50), pipeline(40)
C	1421 (69%)	a name nothing in the tree declares	cast_i8(210), compose(76), booth_mul_i32(64), approximately_equal(62)

Three causes needing three different repairs. B is the interesting one: adder_tree and ternary_gemm are the names of specs, called as if the module were a function — a convention nothing resolves.

A correction to this issue. I wrote earlier that "the missing declaration is the defect" as if it were one thing. cast_i8 is in bucket C and len in bucket A, and those are not the same problem: len is declared somewhere and does not arrive; cast_i8 is declared nowhere at all.

And a larger family this issue did not mention, because nothing had ranked the classes:

642  unexpected type name 'Trit': expected expression
412  use of undeclared identifier 'POS'
319  use of undeclared identifier 'NEG'
───
1373

That is more than twice the __auto_type class three passes went into, and it has never been looked at. It is probably the right next subject.