feat: optional regex engine, per-instance compile cache, testing helpers, and reverse parser (#280)

Ships the opt-in regex-engine backend and the builder-ergonomics polish
end to end.

## Optional regex-engine backend

- Introduces `edify/compile/backend.py` — a small backend layer that
owns the "compile a rendered pattern string with these flags" operation
per engine.
- `edify.RegexBuilder().*.to_regex(engine="re" | "regex")` now actually
routes: `"re"` uses the stdlib :mod:`re` module (default, unchanged);
`"regex"` uses the third-party `regex` module.
- The `regex` module is a **deferred import** — `import edify` never
imports it. When the caller selects `engine="regex"` without the extra
installed, they get a clean, annotated `MissingRegexBackendError`
telling them to `pip install edify[regex]`, not a `ModuleNotFoundError`
deep in a traceback.
- Under `engine="regex"`, **variable-width lookbehind** compiles cleanly
(`assert_behind` bodies with quantifiers like `.between(1,
3).string("foo")` no longer fail). Under `engine="re"`, the compile path
catches stdlib re's cryptic "look-behind requires fixed-width pattern"
and re-raises `VariableWidthLookbehindNotSupportedError` — an annotated
message that names the construct and points the caller at
`engine="regex"`.
- Every match method on `Regex` (`match`, `search`, `fullmatch`,
`findall`, `finditer`, `sub`, `subn`, `split`) accepts a per-call
`timeout=` kwarg. Under `engine="regex"` the value flows through to the
underlying pattern; under `engine="re"` it raises
`TimeoutNotSupportedByEngineError`.
- Two new CI jobs — one installs edify without the extra and verifies
both the friendly ImportError path and the re engine's happy path; the
other installs `edify[regex]` and verifies both engines compile plus the
variable-width-lookbehind fixture only compiles under `engine="regex"`.

## Builder ergonomics

- **Per-instance lazy compile cache.** The first no-kwargs `.to_regex()`
(or match verb) compiles once and caches the resulting `Regex`; every
subsequent no-kwargs call returns the same object (`builder.to_regex()
is builder.to_regex()`). A chain step or `fork()` produces a fresh
builder with its own empty cache. Kwarg calls bypass the cache and
always compile fresh.
- **Match-verb surface closed at five.** `MatcherMixin` now exposes
exactly `test` / `match` / `search` / `findall` / `sub` on the builder.
The four less-common verbs (`fullmatch`, `finditer`, `subn`, `split`)
are reached through `.to_regex()` on the `Regex` wrapper. A discipline
test asserts the closed surface and the absence of every other
`re.Pattern` attribute (`groups`, `groupindex`, `pattern`, `flags`).
- **`Pattern.__call__` delegates to `.test`** — so a `Pattern` doubles
as a validator callable (`email(value) -> bool`) without an intermediate
`.test` step.
- **Testing helpers.** `.assert_matches([...])` and
`.assert_rejects([...])` raise annotated `PatternDidNotMatchInputsError`
/ `PatternMatchedRejectedInputsError` (both subclasses of
`AssertionError` so pytest introspects them exactly like a bare
`assert`).
- **Match wrapper.** Every match verb now returns an edify `Match` that
exposes named captures as attributes — `m.username` and
`m.captures.username` — while delegating everything else (`group()`,
`groupdict()`, `span()`, ...) to the underlying `re.Match` via
`__getattr__`. `re.Pattern.sub`/`subn` callables receive the wrapped
`Match`.
- **`RegexBuilder.from_regex(pattern)` reverse parser.** Converts a raw
regex string back into a builder chain via `re._parser.parse`, covering
the common construct set (literals, character classes with
ranges/categories, quantifiers greedy + lazy including
`?`/`*`/`+`/`{n}`/`{m,n}`, capture / named-capture / non-capturing
groups, all four lookarounds, alternation over literals, anchors,
categories). Anything unrecognized raises the annotated
`UnsupportedReverseParseError`.
- **Per-call allocation win.** Cached `os.path.abspath` for
caller-context filenames, cached `co_positions()` per code object, and
moved `source_line` reading behind a lazy property on `CallerContext`.
Chain-step allocation is ~4.6× faster on a 300-step build/compile
microbenchmark (9.4s → 2.0s / 1000 iterations).
- **Immutability contract documented.** Dedicated section in the API
reference + one-liner in the README explaining that every chain method
returns a new builder, that branching (`base.exactly(2).digit()` vs
`base.end_of_input()`) is always safe, and that repeat `.to_regex()`
returns the cached instance.

## Breaking changes

- `MatcherMixin` no longer exposes `fullmatch`, `finditer`, `subn`,
`split` on the builder — call `.to_regex()` first and use them on the
returned `Regex`.
- `Regex.match` / `.search` / `.fullmatch` and `Regex.finditer` now
return the edify `Match` wrapper (or an iterator of it) rather than a
bare `re.Match`. Existing code that calls `.group(...)`, `.groupdict()`,
`.span(...)`, etc. keeps working via `__getattr__` delegation.
- `EngineNotWiredError` is deleted; both engines are wired now, so the
class had no live callers.

Closes #74, closes #115, closes #116, closes #117, closes #118, closes
#119, closes #120, closes #121, closes #129, closes #138, closes #140,
closes #141, closes #142, closes #143, closes #144, closes #149, closes
#150, closes #151, closes #175.

1 files changed, 113 insertions, 0 deletions

diff --git a/tests/builder/reverse.test.py b/tests/builder/reverse.test.py
new file mode 100644
index 0000000..1ed00ec
--- /dev/null
+++ b/tests/builder/reverse.test.py
@@ -0,0 +1,113 @@
+"""Tests for :meth:`RegexBuilder.from_regex` — the reverse parser."""
+
+import re
+
+import pytest
+
+from edify import RegexBuilder
+from edify.builder.reverse import UnsupportedReverseParseError
+
+
[email protected](
+    ("source", "expected"),
+    [
+        ("hello", "hello"),
+        ("a", "a"),
+        (r"\d+", r"\d+"),
+        (r"\d*", r"\d*"),
+        (r"\d?", r"\d?"),
+        (r"\d{3}", r"\d{3}"),
+        (r"\d{2,5}", r"\d{2,5}"),
+        (r"\d{2,}", r"\d{2,}"),
+        (r"\w", r"\w"),
+        (r"\W", r"\W"),
+        (r"\s", r"\s"),
+        (r"\S", r"\S"),
+        (r"\D", r"\D"),
+        ("[a-z]", "[a-z]"),
+        ("[abc]", "[abc]"),
+        ("^", "^"),
+        ("$", "$"),
+        (r"\b", r"\b"),
+        (r"\B", r"\B"),
+        (".", "."),
+        ("(?:foo)", "foo"),
+        ("(abc)", "(abc)"),
+        ("(?P<name>abc)", "(?P<name>abc)"),
+        ("foo|bar", "(?:foo|bar)"),
+        ("(?=x)y", "(?=x)y"),
+        ("(?!x)y", "(?!x)y"),
+        ("(?<=x)y", "(?<=x)y"),
+        ("(?<!x)y", "(?<!x)y"),
+    ],
+)
+def test_from_regex_translates_common_constructs_faithfully(source, expected):
+    builder = RegexBuilder.from_regex(source)
+    assert builder.to_regex_string() == expected
+
+
[email protected](
+    "source",
+    [
+        "hello",
+        r"\d+",
+        "^abc$",
+        r"[a-z]+",
+        "(?P<x>abc)",
+        "foo|bar",
+        r"(?=x)\w+",
+    ],
+)
+def test_round_trip_compiled_regex_matches_the_same_inputs(source):
+    original = re.compile(source)
+    reversed_builder = RegexBuilder.from_regex(source)
+    reversed_compiled = reversed_builder.to_regex()
+    for candidate in ["", "a", "abc", "abc123", "xyz", "foo bar", "hi"]:
+        assert bool(original.search(candidate)) == bool(reversed_compiled.search(candidate))
+
+
+def test_from_regex_raises_for_unsupported_construct():
+    with pytest.raises(UnsupportedReverseParseError):
+        RegexBuilder.from_regex(r"(a)\1")
+
+
+def test_from_regex_raises_for_alternation_with_non_literal_branch():
+    with pytest.raises(UnsupportedReverseParseError):
+        RegexBuilder.from_regex(r"a|\d+")
+
+
+def test_from_regex_raises_for_unsupported_class_member_shape():
+    with pytest.raises(UnsupportedReverseParseError):
+        RegexBuilder.from_regex(r"[\da-z_]")
+
+
+def test_from_regex_lazy_quantifier_translates_via_lazy_variants():
+    builder = RegexBuilder.from_regex(r"a+?")
+    assert builder.to_regex_string() == "a+?"
+
+
+def test_from_regex_lazy_between_translates_via_lazy_variant():
+    builder = RegexBuilder.from_regex(r"a{2,5}?")
+    assert builder.to_regex_string() == "a{2,5}?"
+
+
+def test_from_regex_lazy_zero_or_more_translates_via_lazy_variant():
+    builder = RegexBuilder.from_regex(r"a*?")
+    assert builder.to_regex_string() == "a*?"
+
+
+def test_from_regex_named_group_reproduces_the_name():
+    builder = RegexBuilder.from_regex("(?P<username>[a-z]+)")
+    emitted = builder.to_regex_string()
+    assert emitted == "(?P<username>[a-z]+)"
+
+
+def test_unsupported_error_message_names_the_offending_construct():
+    with pytest.raises(UnsupportedReverseParseError) as excinfo:
+        RegexBuilder.from_regex(r"(a)\1")
+    assert "hand-write" in str(excinfo.value)
+
+
+def test_from_regex_produces_a_regexbuilder_not_a_pattern():
+    builder = RegexBuilder.from_regex("abc")
+    assert isinstance(builder, RegexBuilder)