gen-c: real Zig builtins (@abs, @intFromEnum, @min, ...) reach the C verbatim -- 85 of 1362 outputs, 769 sites #7297

What happens

t27c gen-c lowers a bare abs(x) to the t27_abs macro (a C11 _Generic over fabs/fabsf/llabs), but it writes the Zig spelling of the same call, @abs(x), into the C as it stands. gcc reads the @ as a stray character and then calls C's own int abs(int):

module Probe;

pub fn d(x: f64) -> f64 {
    return @abs(x);
}

test "abs of a negative double" {
    assert(d(-2.5) == 2.5);
}

$ t27c gen-c probe.t27 > probe.c ; echo $?
0
$ grep -n abs probe.c | head -1
28:    return @abs(x);
$ cc -std=c11 -c probe.c
probe.c:28:12: error: stray '@' in program
probe.c:28:13: warning: implicit declaration of function 'abs' [-Wimplicit-function-declaration]
probe.c:28:17: warning: 'abs' argument 1 type is 'double' where 'int' is expected ... [-Wbuiltin-declaration-mismatch]

Master's t27c writes the same line 28.

How much

Measured on the Railway lab with t27c built from 403b27f + #7242 + #7281 + #7292, running gen-c on all 1363 .t27 files of the #7191 fourth-slice tree (1362 produce output):

• 116 outputs contain a @name( call, 1219 sites in all.
• 85 outputs, 769 sites, call one of the 28 builtins t27c knows (the list is in A call to a builtin that does not exist typechecks clean, generates, and reaches the C verbatim -- 41 specs classed Working do it #4561). Without @setEvalBranchQuota, 59 outputs.
• Per builtin: @setEvalBranchQuota 402, @abs 91, @intFromEnum 81, @floatFromInt 51, @min 33, @bitCast 22, @sqrt 17, @floatCast 11, @exp 11, @max 10, @intFromFloat 8, @enumFromInt 8, @log 6, @round 5, @rem 5, @mod 5, @trunc 1, @floor 1, @ceil 1.
• The other 450 sites name no Zig builtin (@len, @print, @contains, ...). That is A call to a builtin that does not exist typechecks clean, generates, and reaches the C verbatim -- 41 specs classed Working do it #4561's class, not this one.

Why it shows now

#7292 stops the use resolver from reading @abs as a reference to a spec's own abs. Six specs import math::constants whole and call abs only as @abs: ml/activation/gelu_activation, ml/activation/sigmoid_activation, tri/math/bezier, tri/math/constants, tri/math/matrix and tri/math/probability.

• Before use resolver: a builtin call @pow(...) is read as a reference to pow and splices the spec's own pow #7292, math/constants.t27's abs was spliced into each of them as dead code, and its double abs(double) gave gcc's error recovery a prototype for the call after the stray @.
• After it, cc -fsyntax-only reports the same number of errors on all six, but more -Wbuiltin-declaration-mismatch warnings, e.g. probability from 5 to 56.
• None of the six outputs compiles, before or after.

Proposal

Route each sigil builtin that has a C meaning through the lowering its bare spelling already has, as the Rust backend does for both spellings in zig_builtin_to_rust:

• @abs to t27_abs;
• the libm names (@sqrt, @exp, @log, @floor, @ceil, @round, @trunc, @sin, @cos, @tan) to their C functions;
• the casts (@intFromEnum, @floatFromInt, @intFromFloat, @floatCast, @intCast, @as) to a C cast;
• @setEvalBranchQuota, a comptime hint, to nothing.

The emitter lives in bootstrap/src/compiler.rs, which carries FROZEN_HASH.

Acceptance criteria

• The probe above compiles with cc -std=c11 -c with 0 errors and 0 warnings, and a test that checks it fails with the change reverted.
• The number of real-builtin @name( sites in the corpus gen-c output drops from 769, and the new number is stated.
• No output that passed cc -std=c11 -fsyntax-only before fails after.
• Every seal whose gen_hash_c changes is listed.

Boundary

• bootstrap/src/compiler.rs

User Scenarios

• Given a .t27 file that calls @abs(x) on a floating-point value, when t27c gen-c compiles it, then the generated C file contains a call to the t27_abs macro (or equivalent C11 _Generic selection) instead of the literal token sequence @abs(x).
• Given a .t27 file that calls any of the 28 Zig builtins that have a C lowering (e.g., @intFromEnum, @min, @sqrt, @setEvalBranchQuota), when t27c gen-c compiles it, then the generated C file contains the corresponding C construct (cast, function call, macro, or nothing for comptime hints) and no stray @ character.
• Given the full Railway fourth-slice corpus (1362 files that produce output), when t27c gen-c runs on all of them, then the number of real-builtin @name( sites in the combined C output is strictly less than 769, and every file that previously passed cc -std=c11 -fsyntax-only still passes.

Requirements

FR-001: The C emitter in bootstrap/src/compiler.rs MUST lower @abs to the t27_abs macro (C11 _Generic over fabs/fabsf/llabs) for all integer and floating-point argument types.

FR-002: The C emitter MUST lower libm builtins (@sqrt, @exp, @log, @floor, @ceil, @round, @trunc, @sin, @cos, @tan) to their corresponding C math.h functions with correct type suffixes.

FR-003: The C emitter MUST lower cast builtins (@intFromEnum, @floatFromInt, @intFromFloat, @floatCast, @intCast, @as) to explicit C casts.

FR-004: The C emitter MUST lower @min and @max to the t27_min/t27_max macros (or equivalent _Generic selections) matching the bare min/max lowering.

FR-005: The C emitter MUST lower @setEvalBranchQuota to nothing (elide the call entirely) because it is a comptime hint with no runtime semantics.

FR-006: The C emitter MUST apply the same lowering for sigil (@name) and bare (name) spellings of every builtin listed above, consistent with the Rust backend's zig_builtin_to_rust.

FR-007: After the change, the probe example in the issue MUST compile with cc -std=c11 -c producing zero errors and zero warnings.

FR-008: No output file that previously passed cc -std=c11 -fsyntax-only MUST fail after the change.