fix(llm): fail on an empty completion where it happens, and make the cap overridable

[arthapraha]

Description

A completion with no visible content currently travels as None from
LiteLLMBackend into parse_plan, where it reaches a regex and raises:

TypeError: expected string or bytes-like object, got 'NoneType'

The traceback names _TASK_RE.finditer — four frames from the cause — and
mentions neither max_tokens nor finish_reason. Full report in #511.

Fix Details

Two changes, both small.

1. Raise where the fact is. EmptyCompletionError is raised in
LiteLLMBackend.generate_with_usage, after the completion returns and before
LLMResult is constructed — so a None never enters the system. The message
carries the model, finish_reason, tokens consumed and the cap:

<model> returned no content (finish_reason='length'); the completion hit the
token cap of 2048 after 2048 tokens, which reasoning models can consume
entirely on reasoning_content. Raise the cap with AOB_LLM_MAX_TOKENS.

2. Make the cap overridable. max_tokens reads AOB_LLM_MAX_TOKENS,
defaulting to the current 2048 — so behaviour is unchanged unless someone
changes it.

What this deliberately does not do: no retry, no fallback, no silent empty
string. Those are policy decisions for the project; this makes the failure
legible so the policy can be chosen with the facts visible.

Impact on Benchmarking

• No change to baselines: this only converts a crash into a named error.

A run that previously died with a TypeError now dies with an
EmptyCompletionError naming the cause. No run that previously succeeded
behaves differently — the raise fires only where content is None, which is
already fatal today.

Related Issues

• Fixes the defect reported in An empty completion crashes plan_execute with a TypeError that blames the regex #511.

Verification Steps

1. tests/integration does not exist at e11d1c1. What was run instead, from
   src/:

   python -m pytest llm/tests/ -q
   

   main at e11d1c1: 19 passed
   this branch: 25 passed (+6)

   The six new tests, one per behaviour:

   test	what it pins
   empty_content_raises_where_it_happens	content=None, finish_reason='length' → raises, cap named
   empty_content_for_other_reasons_still_raises	not only the length case
   ordinary_completions_are_unaffected	the success path is untouched
   empty_string_is_not_an_empty_completion	"" is a valid answer and must pass through — only None raises
   token_cap_is_overridable	AOB_LLM_MAX_TOKENS takes effect
   token_cap_defaults_to_the_previous_value	absent the env var, still 2048

2. The 6 pre-existing failures in src/evaluation/tests/ are untouched — this
   branch does not import from evaluation.

3. AOB_LLM_MAX_TOKENS is read in-process and never crosses a spawn.
   os.environ.get at litellm.py:49 and litellm.completion at :61 are in
   the same function in the same process, litellm is imported in-process, and
   there is no subprocess or Popen anywhere under src/llm/. Checked
   deliberately: a fix introducing a new environment variable should not depend
   on that variable surviving a process boundary.

Checklist

• I have added tests that prove my fix is effective. (+6)
• My code follows the project's Ruff formatting and linting rules.
  ruff check src/llm/ — all checks passed; ruff format --check src/llm/ — 9 files already formatted, on this branch. Unlike the
  scorer files, this path is clean both before and after.
• I have signed off my commits (DCO).