elegance: type LanceDBBackend.get_by_source() failure instead of swallowing it to []

[Antawari]

DO NOT MERGE — fleet quality burn unit; opened for review only.

The unit

LanceDBBackend.get_by_source() caught except Exception as exc: and returned
[], making a genuine LanceDB lookup failure (corrupt table, unreachable store,
a vector-op error) indistinguishable from "no entries for this source path".
That is the swallowed-failure the Elegance Law forbids — and the exact masquerade
its structural sibling query() already had removed (it now raises a typed
RetrievalError). get_by_source() was the overlooked sibling: same
table.search(...).where(...).to_list() shape, same "list of entries" return,
same indistinguishable [].

It is the deliberate inverse of exists(), whose fail-open-to-False is a
documented, separately-test-pinned dedup contract (a false negative there only
causes a harmless duplicate store). get_by_source() has no such fail-open
caller — its list is the answer — so, like query(), it must raise rather
than degrade. That exists() contract is left untouched.

The falsifiable claim it satisfies

An oracle: inject a table whose .search() raises; get_by_source returns
[] with no exception — refuting any claim that this backend speaks failure
in the one typed vocabulary.

After this change the oracle is satisfied: a thrown lookup surfaces as a typed,
retryable RetrievalError (code == "retrieval", retryable is True) carrying
the originating exception as __cause__ and structured context (source_path).
A genuine empty result (zero matching entries) stays a legitimate success and
still returns [].

TDD

tests/unit/test_knowledge_backend_get_by_source_typed_failure.py (RED first):
4 tests — typed-raise on backend failure, __cause__ chaining, source_path
in context, and zero-hits-is-a-success. Uses a fake exploding table
(lancedb is not a test dependency), mirroring the sibling query() test.

Budget / anti-gaming notes

• The implementation is net-zero in line/statement count
  (backend.py measure 244 → 244), so the src/bonfire file budget is
  unmoved — no budget was inflated.
• The only baseline ratchet is the tests/unit package total re-frozen
  58858 → 59014 (+156, exactly the new test file, which is a new file under
  the 500-line ceiling), the documented re-anchor per
  configs/BASELINE-CONVENTIONS.md. No file row changed; no offender was
  laundered.

Gate

Reproduced locally against a clean export at the pinned kit SHA: CI lint
(ruff check + ruff format --check), full pytest (4421 passed), the
BubbleGum kit battery (cf-sticky-check, cf-file-budget: clean,
cf-mirror-check, cf-recursion-check, cf-exemptions, mypy ratchet =
0 new errors), and the smoke gate (wheel + sdist build/twine/smoke) — all green.