t27b: f32 @log, @log2 and @log10 from libm.t27 (compiler_rt logf, log2f, log10f) #7822

Part of #7746 (t27b passes specs/numeric/gf16.t27), which is part of #6063 (t27b coverage), item C10 of #6488. This is the third PR of #7746's split, after #7791 and #7819.

What needs it

• gf16.t27. gf16_log, gf16_log2 and gf16_log10 call @log, @log2 and @log10 on an f32, and tests such as gf16_log2_eight and gf16_log10_ten reach them. std.math.pow(f32, ..), the last PR of t27b: specs/numeric/gf16.t27 after the type mismatch family -- @bitCast, f32 rounding and libm, std.math, @setEvalBranchQuota #7746, calls @log and @exp of an f32 too.
• specs/ml/loss/kl_divergence.t27 is blocked only on its f32 @log.
• t27b today. It refuses an f32 @log (libm.t27 has no f32 routine for it), and @log2 and @log10 of any type.

What the reference does

Measured on the t27c lab (zig 0.16.0, x86_64 Debug, no libc):

• A run-time @log, @log2 or @log10 of an f32 is a call of compiler_rt logf, log2f or log10f (log.zig, log2.zig, log10.zig). These are musl's fdlibm-style routines on the bits, not the table-driven f64 log that libm.t27 already holds.
• None is correctly rounded, so t27b must return their bits.
• compiler_rt.want_float_exceptions is true on x86_64 and aarch64, so:
  • a zero gives -1 / (x * x) (-inf);
  • a negative gives (x - x) / 0.0, a NaN whose bits are the platform's;
  • +inf and NaN come back as they are, a signalling NaN included.
• A typed f32 operand Zig knows at compile time folds to the same bits as the run-time call. Measured on 24000 points per builtin: the whole bit range, and densely near 1. 0 differ.

Plan

• Port logf, log2f and log10f to t27 in specs/tri/t27b/libm.t27, step for step, with the same constants (as exact decimals of their f32 values) and the same operation order.
• In specs/tri/t27b/libm_plan.t27:
  • an f32 @log calls logf;
  • an f32 @log2 and @log10 call log2f and log10f;
  • an f64 @log2 and @log10 are refused, because compiler_rt's f64 log2 and log10 are not ported.
• Hold each port to Zig's builtin bit for bit in a spec of its own, over sweeps of bit patterns. Then search for inputs where one ulp of a constant shows, as t27b: @log of an f64 calls compiler_rt's log written in t27, from the libm plan (Closes #7217) #7648 did for log.

Acceptance criteria

• t27b test --blockers specs/numeric/gf16.t27 names no @log, @log2 or @log10.
• specs/ml/loss/kl_divergence.t27 no longer stops at ExprCall(@log).
• The new tests pass under t27c test-report with 0 vacuous tests, and under t27b. Mutants of the ports' constants are killed.
• The lab corpus receipt compare gives IMPROVED_ONLY, EQUIVALENT or NEUTRAL.

User Scenarios

• Given gf16.t27 calls @log, @log2, @log10 on f32 values, When t27b compiles gf16.t27, Then the build succeeds without "libm.t27 has no f32 routine for it" errors.
• Given kl_divergence.t27 uses f32 @log, When t27b compiles kl_divergence.t27, Then compilation proceeds past the ExprCall(@log) that currently stops it.
• Given a t27 spec exercising f32 @log/@log2/@log10 over bit-pattern sweeps, When t27c test-report runs it, Then all tests pass with 0 vacuous tests and mutant constants are killed.

Requirements

• FR-001: specs/tri/t27b/libm.t27 MUST contain f32 logf, log2f, log10f ports matching compiler_rt step-for-step with identical constants (as exact decimals of their f32 values) and operation order.
• FR-002: specs/tri/t27b/libm_plan.t27 MUST map f32 @log → logf, f32 @log2 → log2f, f32 @log10 → log10f, and MUST refuse f64 @log2 and @log10.
• FR-003: New spec files for each port MUST verify bit-for-bit agreement with Zig's builtin @log, @log2, @log10 on f32 over full bit-range sweeps and dense sampling near 1.
• FR-004: The ports MUST reproduce compiler_rt exception behavior: zero → -inf, negative → NaN (platform bits), +inf/NaN → unchanged (including signalling NaN).
• FR-005: t27b test --blockers specs/numeric/gf16.t27 MUST report zero blockers for @log, @log2, @log10.

Boundary

• specs/tri/t27b/libm.t27
• specs/tri/t27b/libm_plan.t27
• specs/numeric/gf16.t27
• specs/ml/loss/kl_divergence.t27
• scripts/tri_loop/t27b.py