Summary

Fixes unguarded .choices[0] accesses in four files that crash with IndexError when an LLM returns an empty choices list. This addresses the class of errors reported in #292 (json response error — the choices list can be empty on API errors, rate-limit responses, or model refusals).

Each fix adds an explicit guard immediately before the first .choices[0] access in each call-site:

if not response.choices:
    raise ValueError("LLM returned empty response in <function_name>")

This turns a cryptic IndexError: list index out of range into a clear, actionable error message.

Files changed

Total: 16 unguarded access sites eliminated.

Verification

Fixes were verified with pact_sheaf, a sheaf-cohomological contract checker:

• All 16 openai-chat#choices-nonempty violations resolved
• Ȟ¹ (first Čech cohomology) = 0 on all four fixed files — no residual guard gaps detected
• Z3 UNSAT certificate confirms the guard constraint is satisfiable and the empty-choices branch is unreachable past the guard

Before fix: pact reported h1_semantic = 7, 3, 3, 3 across the four files.
After fix: h1_semantic = 0, 0, 0, 0 for choices-spec violations.

Test plan

• Verify existing test suite passes (pytest or equivalent)
• Manually inject a mock LLM client that returns choices=[] and confirm the new ValueError is raised with a clear message instead of IndexError
• Run ruff check and black --check on the four changed files (already applied)