fix(tensor): relu/leaky_relu/softmax/log_softmax accept a scalar (#632)

[InauguralPhysicist]

Closes #632

Four tensor functions returned a silent null — not a value, not an error — when handed a scalar, while every sibling in the same docs/BUILTINS.md tables (sqrt, exp, log, negative, add) is scalar-clean.

tensor_to_flat returns NULL for a 0-dim VAL_NUM, and each function did if (!flat) return make_null().

Why the silent null is the real problem

It propagates until it hits an unrelated operator, so the failure surfaces far from the cause:

define layer(x) as: return relu of x
layer of ([-3, 4])     # [0, 4]
layer of -3            # null      <- silently poisons
(relu of -3) + 1       # Error line N: cannot apply '+' to null and num

A generic layer that works on a batch quietly poisons on a single sample. Same shape as the #490–#512 silent-tolerance train.

Fix

A scalar fast path at the top of each — the degenerate element-wise case, matching sqrt/exp/log:

function	scalar result
relu of x	max(0, x)
leaky_relu of x	x < 0 ? 0.01x : x
softmax of x	1.0 (one-element normalization)
log_softmax of x	0.0 (log(1))

The list/buffer/2-D paths are untouched — relu of [-3, 4] still gives [0, 4].

Verification

All from the issue, confirmed on main:

relu of -3        -> 0      leaky_relu of -10 -> -0.1
relu of 5         -> 5      softmax of 1      -> 1
relu of [-3, 4]   -> [0,4]  log_softmax of 1  -> 0
(relu of -3) + 1  -> 1      (no more null poison)

Six regression asserts added to test_flat_buffer_tensor.eigs, each failing on the pre-fix null (including the downstream-poison case). BUILTINS.md documents the scalar behavior for all four.

Gates

• Release suite 3078/3078
• ASan + UBSan detect_leaks=1 3082/3082, leak tally 0
• doc-drift clean

🤖 Generated with Claude Code