fix: fail fast when sharing segmenter weights without a compressor

[ClaireXi99]

What breaks

LlmLingua2Segmenter.__init__ has two initialization paths:

if shared is False:
    self.model = AutoModel.from_pretrained(...)      # loads its own weights
    self.tokenizer = ...
    self.buffer_len = ...
elif compressor is not None:
    self.model = compressor.inner_compressor.model   # shares the compressor's weights
    self.tokenizer = ...
    self.buffer_len = ...

When shared is True and no compressor is supplied, neither branch runs. The object is constructed successfully but without model, tokenizer or buffer_len, so the failure surfaces later and far from its cause — LightMemory.__init__ immediately reads self.segmenter.buffer_len, giving:

AttributeError: 'LlmLingua2Segmenter' object has no attribute 'buffer_len'

That message points at the segmenter's internals rather than at the actual problem, which is a configuration one: precomp_topic_shared=True asks the segmenter to share a pre-compressor's weights while pre_compress=False means there is no pre-compressor to share.

Reproduced directly against the class:

>>> seg = LlmLingua2Segmenter(config={"model_name": "dummy"}, shared=True, compressor=None)
>>> [hasattr(seg, a) for a in ("model", "tokenizer", "buffer_len")]
[False, False, False]
>>> seg.buffer_len
AttributeError: 'LlmLingua2Segmenter' object has no attribute 'buffer_len'

Fix

Add the missing else branch so an unsatisfiable configuration fails immediately with an actionable message naming both ways out:

precomp_topic_shared=True requires a pre-compressor to share weights with, but none
was provided. Either set pre_compress=True and configure pre_compressor, or set
precomp_topic_shared=False to let the segmenter load its own model.

This is intentionally narrow — it only converts a silent half-initialization into a clear error. Neither existing branch changes behavior.

Relationship to PR #75

PR #75 fixes a related crash one layer up, where LightMemory passed an unset self.compressor to the segmenter factory. The two changes are complementary rather than overlapping: #75 touches only lightmem.py, this one only llmlingua_2.py, and I have deliberately left #75's one-line fix out of this branch to avoid duplicating it.

Worth noting that #75 makes this bug more reachable. Once self.compressor correctly defaults to None, the precomp_topic_shared=True plus pre_compress=False combination stops failing on the missing attribute and starts flowing into the segmenter as compressor=None with shared=True — landing exactly on the branchless path above. That is the same cascade described in PR #63, where PR #59's fix removed an early return and exposed an uninitialized variable underneath.

With precomp_topic_shared at its default of False nothing changes: shared is False takes the first branch and the segmenter loads its own model, so #75's scenario and its assertion of (topic_segmenter, False, None) are unaffected.

Tests

Added tests/test_topic_segmenter_shared_weights.py covering both sharing paths:

• shared=True with no compressor raises the new ValueError
• shared=True with a compressor still adopts its model, tokenizer and max_position_embeddings-derived buffer_len

The shared=False path is unchanged and untested here because it downloads a real model. The module imports torch, so the file uses pytest.importorskip("torch").

$ python -m pytest tests/ -v
tests/test_sensory_memory.py::test_oversized_single_user_message_is_consumed PASSED
tests/test_sensory_memory.py::test_force_segment_flushes_remaining_buffer PASSED
tests/test_topic_segmenter_shared_weights.py::test_shared_without_compressor_raises_clear_error PASSED
tests/test_topic_segmenter_shared_weights.py::test_shared_with_compressor_reuses_its_weights PASSED

4 passed

Verified on Python 3.11 with the pinned transformers==4.57.0; the pre-existing sensory-memory tests still pass.