fix(cartesia,pipecat): fall through to chunk when memory is empty

[Agnik47]

Stacked on #1683. This branch has that PR's commit as its parent, so the diff below shows both. Only the second commit belongs to this PR: 3d19805. Happy to rebase onto main once #1683 lands.

Summary

_field() returns the first value that is not None. A v4 hybrid-search hit shaped {"memory": "", "chunk": "..."} therefore resolves to the empty string and never falls through to chunk — the exact fallback the call site's own comment says it is there for:

# v4 search.memories/hybrid uses `memory` or `chunk`.
memory = _field(r, "memory", "chunk", "content", default="")

It bites on both paths:

• deduplicate_memories — the empty text produces an empty comparison key, so the hit is dropped and the memory never reaches the prompt.
• format_memories_to_text — when such an item does survive, the same _field call renders it as a bare - [16 hrs ago] .

Why this is the odd one out

The TypeScript getMemoryText in packages/tools/src/tools-shared.ts takes the first non-empty field. I differential-tested all four Python SDKs against it:

input	TS (reference)	agent-framework	openai	cartesia	pipecat
{"memory": "", "chunk": "text"}	['text']	['text']	['text']	[]	[]
{"memory": " ", "chunk": "text"}	['text']	['text']	['text']	[]	[]
SimpleNamespace(memory="", chunk="text")	['text']	['text']	['text']	[]	[]
{"memory": None, "chunk": "text"}	['text']	['text']	['text']	['text']	['text']
{"chunk": "text"}	['text']	['text']	['text']	['text']	['text']
{"memory": "mem", "chunk": "chk"}	['mem']	['mem']	['mem']	['mem']	['mem']

agent-framework-python and openai-sdk-python already match the reference — they filter on isinstance(value, str) and value.strip(). The two voice SDKs were the only implementations that disagreed, which is why #1683 (a cartesia↔pipecat parity fix) did not surface it.

Changes

• _memory_text() in both voice SDKs, mirroring the TypeScript getMemoryText semantics, used on the dedup and the render path. Keeping both on one helper is what stops a rescued hit from rendering as an empty bullet.
• packages/pipecat-sdk-python/tests/conftest.py — pipecat's import stubs lived inside test_empty_profile.py, so a second test module could only import the package when collection happened to run that file first. Moving them to conftest.py makes tests/test_utils.py importable on its own. test_empty_profile.py is untouched; its own guarded copy becomes a no-op.

Test plan

• 4 new tests per package fail without the source change and pass with it:

  FAILED tests/test_utils.py::TestSearchResultExtraction::test_falls_through_to_chunk_when_memory_is_empty
  FAILED tests/test_utils.py::TestSearchResultExtraction::test_falls_through_to_chunk_when_memory_is_whitespace
  FAILED tests/test_utils.py::TestSearchResultExtraction::test_falls_through_on_models_too
  FAILED tests/test_utils.py::TestSearchResultExtraction::test_never_renders_an_empty_bullet
  

• With the fix: cartesia 14 passed, pipecat 8 passed. Pipecat verified isolated and in reverse collection order.
• Differential harness re-run after the change: divergences from the TypeScript reference 0, cartesia↔pipecat behavioural differences 0 across 288 dedup/format input combinations.
• test_memory_still_wins_over_chunk and test_entry_with_no_usable_text_is_still_dropped pin the behaviour that should not change.