fix(llm): raise EmptyCompletionError instead of crashing on empty completions

[virajsabhaya23]

Description

An empty model completion took the whole run down with a TypeError that blamed
a regex four frames from the cause. LLMResult.text was set to None and the
failure only surfaced in parse_plan at _TASK_RE.finditer(raw), with no
mention of max_tokens, finish_reason, or the empty completion.

Fix Details

Reasoning models charge hidden reasoning_content against the same completion
cap, so a demanding question can exhaust a 2048-token budget and return
content: null with finish_reason: "length" regardless of prompt size. The
run reported in #511 had a 256-token prompt and hit the 2048 cap exactly.

• Validate the completion where it arrives, before LLMResult is constructed,
  so a None never enters the system. EmptyCompletionError names the model,
  finish_reason, tokens consumed and the cap.
• Make the cap overridable via AOB_LLM_MAX_TOKENS, defaulting to the current
  2048 so nothing changes unless it is set. Documented in INSTRUCTIONS.md.
• Applied to both LiteLLMBackend and OpenAICompatBackend through a shared
  result_from_response helper. The issue named only the litellm path, but
  openai_compat had the identical unchecked .content read and the same
  hardcoded cap.

Scope note: no retry/backoff here. #511 distinguishes a null completion
(possibly retryable) from a 429 (back off); that is a policy decision worth its
own PR.

Impact on Benchmarking

• No change to baselines: This fix only improves stability/performance.
• Baseline change

Previously-passing runs are unaffected: the default cap is unchanged at 2048 and
a non-empty completion follows the same path. Runs that previously crashed now
raise a named, diagnostic error instead of a misleading TypeError.

Related Issues

• Fixes: An empty completion crashes plan_execute with a TypeError that blames the regex #511

Verification Steps

1. uv run pytest src/llm/tests — 19 passing on main, 34 on this branch.
2. uv run ruff format --check src/llm && uv run ruff check src/llm — clean.
3. New tests cover None / empty / whitespace-only content, the recorded
   256-prompt/2048-completion failure, AOB_LLM_MAX_TOKENS reaching the
   request, and rejection of invalid override values.

Note: the template's uv run pytest tests/integration path does not exist in
this repo; tests live under src/*/tests.

Checklist

• I have added tests that prove my fix is effective.
• My code follows the project's Ruff formatting and linting rules.
• I have signed off my commits (DCO).