[thorwhalen]

Fixes #6, fixes #7.

• insert: exact (duplicate-safe) via WATCH/MULTI with a sentinel + LINSERT; list.insert clamping semantics.
• __delitem__: atomic MULTI of LSET sentinel + LREM; clear() is a single DEL. So pop/remove/reverse/+= now work.
• Slices follow slice.indices (l[:], l[1:0], negatives, steps). Out-of-range int get/set/del raise IndexError (so index()/in terminate). Slice assign/delete: NotImplementedError.
• Iteration in batches of iter_batch_size (one lrange per batch).
• store[k] = [] raises ValueError before touching the existing value (no more data loss); list replace is atomic (DEL+RPUSH in MULTI); generators accepted; numbers are stored as strings instead of being silently dropped.
• 37 tests on fakeredis.

[thorwhalen]

Adversarial review (fresh Opus subagent): no correctness/data-loss bugs. 12,000 randomized differential runs against a python list (raw, decode_responses, codec, tiny batch), plus retry/WATCH, 8-thread concurrency and failure-atomicity probes all passed. Minor findings: non-integer index not TypeError and iter_batch_size<=0 looping forever -> fixed in the last commit (38 tests); wrong-type-key message and bytearray storage -> deferred to a new issue. Dependents gate: redisdol has no dependent suites (see PR #9 note).