feat: regex introspection API — explain, verbose export, ASCII + SVG visualize + annotated errors (#272)

Ships the regex introspection API on the `Regex` wrapper — three new
methods that turn a compiled pattern into something a human can read,
plus a full retrofit of every error message to a shape that shows the
caller *where* the failure happened and *how* to fix it.

## `Regex.explain()`

Plain-English bullet list of what the pattern accepts, followed by a
small set of concrete accepted strings.

```
- The text must start with either "http" or "https".
- Then the text must have "://".
- Then the text must have one or more letters, digits, or underscores.

Text this pattern accepts:
    http://ab1
    https://b1_c
    http://1_c2d
```

## `Regex.to_verbose_string()`

An `re.VERBOSE`-compatible export where every fragment is annotated
inline. Copy-pasting the output into `re.compile(..., re.VERBOSE)`
compiles to the same pattern.

```
(?P<year>               # begin group named "year"
  \d{4}                 # exactly 4
)                       # end group named "year"
```

## `Regex.visualize(format='ascii')`

ASCII railroad diagram — `START` and `END` boxes flanking the pattern,
arrows between elements, fork/merge junctions for alternation,
dashed-caption blocks for captures and lookarounds.

```
                        +--------+
                   +--->| "cat"  |----+
                   |    +--------+    |
                   |                  |
   +-------+       |    +--------+    |   +-----+
   | START |------>+--->| "dog"  |----+-->| END |
   +-------+       |    +--------+    |   +-----+
                   |                  |
                   |    +--------+    |
                   +--->| "fish" |----+
                        +--------+
```

## `Regex.visualize(format='svg', engine='graphviz')`

SVG rendered by Graphviz. Junction-point fork/merge for alternation,
dashed rounded clusters for captures/lookarounds, folded two-line node
labels for quantifiers (e.g. `digit\n(one or more)`). Requires the
optional `graphviz` extra; the missing-dependency error names the
install command.

## Annotated error messages

Every exception across `edify.errors` — anchors, captures, input,
internal, naming, quantifier, structure — now emits a message that names
the offending call in the caller's source, explains the invariant that
was violated, and prescribes the fix:

```
error: start_of_input has already been added to this pattern

  --> user_code.py:14:26
   |
14 |     pattern = RegexBuilder().start_of_input().digit().start_of_input()
   |                                                       ^^^^^^^^^^^^^^^^ second start_of_input added here
   |
   = note: a pattern can carry at most one start_of_input anchor; the earlier .start_of_input() call already set it.

help: remove the duplicate .start_of_input() call from the chain.
```

Caller source location is captured via `sys._getframe()` walking to the
first non-edify frame and `co_positions()` for precise column spans.

## Other landings

- `Regex` retains the AST elements the builder produced; every
introspection method operates on the AST directly rather than parsing
the emitted regex string.
- Lazy compile cache on `BuilderCore` — a builder's `to_regex()` result
is memoised so `.match()` / `.search()` / `.explain()` share one
compiled pattern.
- `Regex.__getattr__` delegates any attribute not on the wrapper to the
underlying `re.Pattern`, so `.pattern`, `.flags`, `.groups`,
`.groupindex` continue to work.
- `Regex.__eq__` / `Regex.__hash__` raise when called on an unfinished
builder, pointing at both the compare/hash call site and the still-open
frame or dangling quantifier.
- Property test extended: every recursively-built composition of
quantifiers, groups, captures, named captures, and subexpressions emits
its expected regex exactly, no fragment dropped or duplicated.

1 files changed, 69 insertions, 0 deletions

diff --git a/tests/builder/cache.test.py b/tests/builder/cache.test.py
new file mode 100644
index 0000000..80862d3
--- /dev/null
+++ b/tests/builder/cache.test.py
@@ -0,0 +1,69 @@
+"""Tests for the lazy :class:`Regex` cache on :class:`BuilderCore`."""
+
+from edify import Pattern, RegexBuilder
+
+
+def test_repeated_lazy_regex_calls_return_the_same_instance():
+    builder = RegexBuilder().digit()
+    first = builder._lazy_regex()
+    second = builder._lazy_regex()
+    third = builder._lazy_regex()
+    assert first is second
+    assert second is third
+
+
+def test_repeated_matcher_calls_reuse_the_cached_regex():
+    builder = RegexBuilder().digit()
+    builder.match("1")
+    cached = builder._cached_regex
+    builder.search("2")
+    builder.findall("3")
+    builder.test("4")
+    assert builder._cached_regex is cached
+
+
+def test_pattern_lazy_regex_is_memoised_too():
+    pattern = Pattern().word()
+    first = pattern._lazy_regex()
+    second = pattern._lazy_regex()
+    assert first is second
+
+
+def test_a_forked_builder_gets_a_fresh_cache_slot():
+    original = RegexBuilder().digit()
+    original.match("1")
+    assert original._cached_regex is not None
+    forked = original.fork()
+    assert forked._cached_regex is None
+
+
+def test_a_copied_builder_gets_a_fresh_cache_slot():
+    original = RegexBuilder().digit()
+    original.match("1")
+    copied = original.copy()
+    assert copied._cached_regex is None
+
+
+def test_a_chain_extension_gets_a_fresh_cache_slot():
+    original = RegexBuilder().digit()
+    original.match("1")
+    extended = original.word()
+    assert extended._cached_regex is None
+
+
+def test_a_fresh_builder_starts_without_a_cached_regex():
+    builder = RegexBuilder()
+    assert builder._cached_regex is None
+
+
+def test_calling_to_regex_directly_does_not_populate_the_cache():
+    builder = RegexBuilder().digit()
+    _ = builder.to_regex()
+    assert builder._cached_regex is None
+
+
+def test_lazy_regex_produces_a_regex_that_matches_the_pattern():
+    builder = RegexBuilder().digit()
+    result = builder._lazy_regex()
+    assert result.source == "\\d"
+    assert result.match("7") is not None