Quoted string keys break extract/Delta; escaping approach needs compatibility direction #648

On current dev (59acdd9), a dictionary key containing both single and double quotes produces a path that cannot be consumed by parse_path, extract, or Delta.

from deepdiff import DeepDiff, Delta, extract

key = '''Bob's "plan"'''
before, after = {key: 1}, {key: 2}
diff = DeepDiff(before, after)
path = next(iter(diff['values_changed']))
print(path)
extract(after, path)  # KeyError
# Independently, before + Delta(diff, bidirectional=True, raise_errors=True)
# raises DeltaError instead of producing after.

Expected: the returned path identifies the original key and the diff can be replayed in both directions. The same failure occurs for added/removed quoted keys. This was reproduced on Windows / CPython 3.12.14.

I prepared a prototype and regression tests using explicit u-prefixed Python string literals only for keys containing both quote types. This separates escaped literals from existing raw-backslash path semantics, including nested keys ending in a backslash. A canonical str.__repr__ call also handles string subclasses.

The prototype has a compatibility trade-off that needs your direction: old malformed paths still worked as opaque exclude_paths strings. Changing their canonical spelling means persisted old exclusions stop matching, while exclusions regenerated from the new output work. Ordinary keys, single-quote-type keys and old raw-backslash paths retain their existing spelling and parsing.

Would a canonical escaped representation with migration guidance be acceptable, or should the solution preserve old opaque exclusion strings as well? I have not opened a PR because this compatibility question is unresolved.

Validation of the prototype:

• Unchanged source: 23 new regression failures; fixed related path/model/glob/search/Delta/text tests: 577 passed, 3 skipped.
• Broader declared-dependency suite including slow tests: 1340 passed, 8 skipped, 2 deselected, 95% source coverage. Security payload tests and restricted-unpickler payload cases were not run; no complete Python/OS matrix is claimed.
• 4,680 generated short-key cases and 3,544 old-valid single paths / 3,721 old-valid nested paths passed, plus independent public API, persisted-Delta and input-preservation checks.
• Pyright and blocking Flake8 checks pass; independent agent review found no remaining implementation issue after the string-subclass fix.

Existing include_paths assembly also has a both-quote limitation on unchanged source and remains outside this prototype; I am not claiming every path consumer is fixed. #640 handles byte keys, while #351/#430 concern earlier quoting cases; this report covers ordinary string keys containing both quote types.

AI assistance: OpenAI Codex investigated and implemented this user-authorized prototype, ran the checks and obtained an independent agent review. No separate human review or personal production incident is claimed.