Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@tests/test_runner.py`:
- Around line 109-124: Tests in TestCreateClient depend on external environment
state; make them deterministic by using pytest's monkeypatch to clear provider
API env vars before calling create_client. Update test_anthropic_prefix,
test_openai_prefix, and test_o1_prefix to accept a monkeypatch fixture and call
monkeypatch.delenv("ANTHROPIC_API_KEY", raising=False) /
monkeypatch.delenv("OPENAI_API_KEY", raising=False) /
monkeypatch.delenv("O1_API_KEY", raising=False) respectively (or the actual env
var names used by create_client), then run the existing with
pytest.raises((ImportError, EnvironmentError)): create_client(...) assertion so
the test no longer relies on CI secrets; keep the
test_unknown_raises_value_error unchanged.
- Around line 206-217: The test_jsonl_round_trip creates a temporary file with
tempfile.NamedTemporaryFile and calls path.unlink() after assertions, but that
cleanup won't run if an assertion fails; update the test to use pytest's
tmp_path fixture or a try/finally so the temp file is always removed.
Specifically, replace the tempfile.NamedTemporaryFile usage in
test_jsonl_round_trip (and the path variable) with tmp_path.joinpath / tmp_path
/ tmp_path fixture APIs to create/write the .jsonl file, or wrap the current
creation/assertion in try/finally to call path.unlink() in the finally block so
cleanup always occurs.

In `@vera_bench/cli.py`:
- Around line 143-144: The code is re-serialising ProblemResult objects via
to_jsonl() and json.loads(), which is wasteful; update compute_metrics (and its
callers like where compute_metrics is invoked before _print_metrics) to accept a
list of ProblemResult objects directly (or alternatively convert each
ProblemResult to a dict with dataclasses.asdict(result) or a dedicated to_dict()
method) and pass results returned from run_benchmark straight into
compute_metrics (replace json.loads(r.to_jsonl()) with either r or asdict(r));
adjust compute_metrics parameter type and internal handling to read fields from
ProblemResult instead of expecting pre-parsed dicts.
- Around line 62-67: The CLI accepts --max-tokens but it isn't forwarded to the
LLM call; update the call chain to thread max_tokens from the click handler into
run(), then into run_benchmark(), then into run_single_problem(), and finally
pass it to client.complete() (or the client's request payload) so the runtime
uses the user-specified value; update the function signatures for run(),
run_benchmark(), and run_single_problem() to accept a max_tokens:int (with
existing defaults preserved) and propagate that parameter when invoking
client.complete().

In `@vera_bench/metrics.py`:
- Around line 68-108: The logic that tallies check/verify/fix/run counts is
duplicated between compute_metrics and _compute_by_tier; extract it into a new
helper _compute_counts(by_problem: dict[str, list[dict]]) that returns the tuple
(check_pass_count, verify_pass_count, verify_eligible, fix_success,
fix_eligible, run_correct_count, run_eligible, total) using the exact selection
logic (attempt_1, attempt_2, best) shown in the diff, then replace the local
counting blocks in compute_metrics and _compute_by_tier to call _compute_counts
and map the returned values into their BenchmarkMetrics constructions (update
the arguments to _rate calls accordingly) so both functions reuse the single
implementation and remain consistent.

In `@vera_bench/models.py`:
- Around line 138-146: The code assumes choice.message is non-null when
computing text (choice.message.content), which can raise AttributeError; modify
the extraction to defensively check that choice and choice.message exist before
accessing .content (e.g., set text = choice.message.content if choice and
choice.message and choice.message.content else ""), update the logic around
response.choices and the LLMResponse construction (references: response.choices,
choice, choice.message, LLMResponse) so text falls back to an empty string when
message is None while preserving the existing usage and model fields.
- Around line 114-147: The complete method is passing timeout=timeout into
self._client.chat.completions.create which the OpenAI SDK 1.x does not accept;
remove the timeout kwarg from that call and instead either instantiate the
client with a timeout or call
self._client.with_options(timeout=timeout).chat.completions.create(...); update
the call site in complete (and any similar calls) to use
client.with_options(timeout=timeout).chat.completions.create(...) or ensure the
client was created with OpenAI(timeout=...) so you avoid the TypeError at
runtime.

In `@vera_bench/runner.py`:
- Around line 27-41: The regex _FENCE_RE used by extract_vera_code requires a
newline before the closing backticks so blocks like ```vera\ncode``` are missed;
update _FENCE_RE to allow an optional newline before the closing backticks (e.g.
make the pattern use \n? before ```), keep re.DOTALL, then ensure
extract_vera_code continues to pick the longest match and returns the stripped
code plus a terminating newline.

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@README.md`:
- Around line 53-74: Add a new "Results summary table" section to README.md that
explains that running `vera-bench report results/` produces `results/summary.md`
and show the expected per-model columns (Model, check_rate, verify_rate,
fix_rate, run_correct_rate, wall_time_s) with a small example row; place this
section near the usage/CLI examples so it satisfies the README requirement to
document the results summary and reference `summary.md` as the source of the
table.