Guard against empty/None/truncated provider responses in rto, self_consistency, reread, leap

[SuperMarioYL]

Summary

Several optimization approaches access response.choices[0].message.content
directly without validating the provider response first. When an upstream
provider returns an empty choices list, a None content, or a
length-truncated completion, this raises an uncaught IndexError /
TypeError and aborts the request with an opaque error.

This extends the response-validation idiom already merged in moa.py,
bon.py, plansearch.py and mcts.py to the four remaining approach modules:

File	Before	After
rto.py	4 raw .choices[0] accesses in the sequential C1→Q2→C2→C3 pipeline; empty choices → IndexError, None content corrupts the next prompt	each step validated, raises an informative error
self_consistency.py	None content appended to the sample list then fed to SequenceMatcher in calculate_similarity → TypeError	empty/None/truncated samples are skipped; degrades to the existing "No consistent answer found." fallback
reread.py	None.strip() / empty choices crash; multi-choice path keeps Nones	validated; multi-choice path keeps only usable, non-truncated choices
leap.py	5 raw accesses, several feeding extract_output / .split() → TypeError on None	validated via a small _extract_content helper at all five call sites

The guard mirrors the one introduced in #266:

if (response is None or
        not response.choices or
        response.choices[0].message.content is None or
        response.choices[0].finish_reason == "length"):
    ...

Single-response steps (rto / reread n=1 / leap) raise an informative error
because they have no alternative to fall back on; the in-loop sampler
(self_consistency) skips the bad sample and continues, matching the
continue behavior in bon/moa.

Behavior note

For rto, reread and leap, a finish_reason == "length" response was
previously returned as a (partial) success and is now treated as a failure.
This is intentional: a truncated intermediate result silently corrupts every
downstream step of these pipelines, so failing loudly is preferable to
returning unreliable output. self_consistency simply drops the truncated
sample.

Scope

Tightly scoped to the four approach modules above plus tests
(~65 functional LoC). It deliberately does not touch the files in #310
(mars/agent, coc_plugin, deep_research, spl) — the two PRs are
complementary and cover disjoint files.

Testing

Added tests/test_approaches.py::test_approaches_handle_bad_responses, which
drives all four approaches with a MockBadClient for the empty-choices,
None-content and length-truncated cases. The test is red on main (uncaught
IndexError from round_trip_optimization on empty choices) and green
with this change. The existing approach tests continue to pass
(finish_reason="stop" was added to the shared MockClient so the happy path
exercises the new guard).

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}