[John-Cusack]

Leaner generated validation code (~1.5x on object-heavy schemas), no behavior change

What

Three pure-Python codegen-shape changes to the generated validate
functions (fastjsonschema/draft04.py only — same shapes apply to
draft06/draft07 if you want them; kept this PR to draft04 to
stay reviewable):

1. required without set ops (generate_required): emit
   [key for key in [...] if key not in data] instead of
   set([...]) - data.keys(). Same exception type, message
   (str(sorted(missing))), definition, and rule; the set
   construction moves off the valid path.
2. no dead _keys set (generate_properties /
   generate_pattern_properties / generate_additional_properties):
   skip building, probing (in), and pruning (.remove) the
   {variable}_keys set when no consumer exists at that object level
   (additionalProperties absent/True/{} and no
   propertyNames: false); direct if "key" in {variable} guards
   instead.
3. enum tuple membership (generate_enum): for all-string enums
   emit if x not in ('a', 'b', ...) instead of the per-option
   isinstance-and-== or-chain. Same message, definition, rule.

Correctness guard included: trivially-checkless schemas
(properties: {}, patternProperties: {}, vacuous
additionalProperties: true/{}) previously relied on the
unconditional _keys line to keep the generated if block non-empty;
with it gone, those shapes emitted an empty if body
(IndentationError at compile). The trivial returns are now hoisted
above the block so no empty if is ever generated (verified: those
schemas compile and validate correctly on this branch).

Measured

Ryzen 9 5950X, CPython 3.12.3, frozen 3000-doc nested-order batch
(compile-once excluded by construction), each arm a fresh subprocess,
in-process timing after warmup, min-of-7 per arm with arms interleaved
round-robin (shared box), peak RSS via /proc VmHWM:

Track	Stock (master @ 738da70)	Patched	Speedup
rep (half-invalid)	0.3355 s	0.2262 s	1.48x
valid-only	0.2942 s	0.1697 s	1.73x
x1 (single small doc)	0.0142 s	0.0074 s	1.92x

Ratios stable across two full interleaved runs (rep 1.48x both runs;
valid 1.62–1.73x; x1 1.88–1.92x). Peak RSS flat across arms (~23 MB
rep/valid, ~17.6 MB x1 both sides). Outcome checksums (sha256 over
valid/invalid outcomes) and error samples (message, name, rule,
definition on the first 40 docs) agree stock-vs-patched on all
tracks.

Reproduce (at 738da70 vs this branch, venv per Makefile
pip install -e .[devel] + the JSON-Schema-Test-Suite submodule):

python bench_fjs.py rep   # + valid, x1; repeat per arm, take min

(Bench scripts with the frozen corpus available on request — say the
word and I'll attach them.)

Test / lint proof

• pytest -W default --benchmark-skip (their make test gate):
  4539 passed, 12 skipped, 435 xfailed, 950 xpassed — zero
  failures (includes the JSON-Schema-Test-Suite submodule cases).
• pylint fastjsonschema: 9.90/10, byte-identical score and warning
  set to unmodified master (no new warnings from this change).
• CHANGELOG.txt entry added under === UNRELEASED, matching
  contributor practice.

[John-Cusack]

Closing this. The title and changelog say no behavior change, which isn't true: {"properties": []} now compiles instead of raising AttributeError, and an all-string enum now accepts a non-string object whose __eq__ matches one of the strings. On the repo's performance.py schema it is also only about 6% faster.