v1.0.3

A performance release. No API changes; results are identical to 1.0.2.

Performance

Profile-driven tuning of the core step evaluator (evaluateStep). Earlier
guesses aimed at cutting object/GC allocation, but on this engine GC is only
~3.5% of runtime — so a CPU profile (node --prof bench/bench.js) was used to
find the real hot spots:

• Megamorphic axis lookup — the axis function is now resolved once per
  step (resolveAxis) instead of once per input node, eliminating a
  megamorphic AXES[axis] keyed-load IC (~6.5% → ~1.6% of ticks).
• Redundant predicate classification — a predicate's "is it a pure node-set
  (existence filter) vs a position test" is a static property of the AST, now
  computed once per step instead of once per input node (isPureNodeSet
  ~3.3% → ~0.1%); predicate-less steps skip the predicate pass entirely.
• Defensive array copies — the child and attribute axes no longer
  .slice() the adapter's arrays. evaluateStep treats axis results as
  read-only (it copies matches into a fresh array), so an axis array is never
  mutated or used as a node-set's backing store. This removes ~5% of slice
  (CloneFastJSArray) overhead; a regression test guards the invariant.
• Fused descendant walk — a selective //name applies its node test during
  the descendant traversal instead of materializing the whole descendant set
  and then filtering it.
• Empty input node-sets short-circuit before any per-step setup.

On the capybara-simulated–shaped benchmark this is ~10–29% faster across the
query mix, with scaling still linear:

query	1.0.2	1.0.3
count all elements	0.160ms	0.114ms
all fields (union-in-predicate)	0.396ms	0.306ms
field by label (//label hoist)	0.743ms	0.597ms
deep descendant attr	0.261ms	0.215ms
link by text	0.281ms	0.231ms
table cell by text	0.570ms	0.512ms

Full Changelog: v1.0.2...v1.0.3