fix: stop silently dropping a failed FAISS reconstruction, and bound the linter

[abrichr]

ruff>=0.1.0 with no explicit [tool.ruff.lint] select means ruff check
here answers differently depending on the day it runs. ruff 0.16.0 grew its
DEFAULT rule set from 59 rules to 413 and this project selected no rules, so it
silently inherited 354 of them. Verified against this tree:

ruff 0.14.x All checks passed!
ruff 0.16.0 Found 103 errors.

This repo has no test/lint CI workflow, so nothing would have caught it; the
only consumer of the linter here is a developer running it locally, which is
exactly the case an unbounded range makes unreproducible.

Real defect (1):

• VectorIndex.get_vectors caught every exception from
  faiss.reconstruct_n(...) and passed, falling through to return None.
  None is this method's DOCUMENTED answer for "FAISS-only storage, nothing to
  hand back", so a failed reconstruction was indistinguishable from the normal
  empty case and a caller would conclude the index simply had no vectors. It now
  catches what FAISS actually raises (RuntimeError, AttributeError) and logs
  a warning naming the vector count and the exception type. The return value is
  unchanged for every input that previously succeeded.

Fixed in code (no behaviour change): UP045/UP007/UP006/UP035 (76) Optional[X]
and Union[...] to PEP 604, Type/Tuple to builtins -- all valid at the
declared requires-python = ">=3.10"; I001 (9) import ordering; F401 (5) unused
imports; PIE790 (5) pass after a docstring; RUF022 (2) __all__ sorting;
SIM102 (2) two nested ifs collapsed in _compute_context_bonus; SIM117 (1)
nested with collapsed in a test.

Disabled deliberately (1, narrowly):

• RUF022 is off for src/openadapt_retrieval/__init__.py only, via
  per-file-ignores with the reasoning inline. That __all__ is grouped by
  subsystem with # Embedders / # Retriever / # Storage headings, and an
  alphabetical sort strands all three comments against unrelated names. The
  grouping documents the package layout and is worth more than the ordering.
  Every other __all__ in the package was sorted.

No select was added. After the above the package clears the whole 413-rule
default set, so narrowing it would discard coverage it is earning today --
including the S110 rule that surfaced the defect. Predictability comes from
bounding ruff>=0.16,<0.17 instead: ruff only grows its default set in a minor
release, so the rule set cannot move until a person raises that ceiling and owns
the new findings in the same change. target-version = "py310" is now stated
rather than inferred.

uv.lock is relocked so ruff actually resolves 0.16.0; without that the bound is
inert for anyone running uv sync. The lock rewrite is large because uv also
upgraded the lockfile format (revision 1 to 3, adding upload-time metadata);
no dependency version changed except ruff, and the local package version was
corrected from a stale 0.1.0 to the declared 0.1.2.

28 passed, 1 skipped (3 gpu/slow deselected, as before). uv run ruff check .
is clean under the locked ruff 0.16.0.

Note: src/**/__pycache__/*.pyc and tests/__pycache__/*.pyc are committed to
this repo and were left untouched here. They are stale build artifacts that do
not belong in version control and should be removed separately.

🤖 Generated with Claude Code

Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01NyCHrzA1psrKMFfroYbzaM