fix(helpers): correct LLM-output parsing (post-v0.0.1)

[will-fawcett]

Summary

Fixes four real correctness bugs in `src/eve_mcp/server/helpers.py` surfaced by the post-v0.0.1 code review. All concentrated in the LLM-output-parsing layer used by `extract_factuality_issues` and `assess_factuality_issue`. Smoke tests against EVE didn't catch these because the existing tests only exercised the happy path with synthetic well-formed inputs.

1. `extract_tag` was greedy — captured across fences

```python
pattern = rf"```{tag}(.*)```" # greedy
```

With input containing two ```json fences, the regex captured everything from the first opening fence to the last closing fence — yielding a string with intervening prose and other fenced blocks embedded.

Fix: `(.*?)` (non-greedy). Regression test added.

2. `_assess_factuality_issue` answer cleanup left tags in the answer

The original code:
```python
code_recommendations = extract_xml_tag(r["answer"], "CODE_RECOMMENDATIONS")
if "```json" in code_recommendations:
code_recommendations = extract_tag(code_recommendations, "json")
r["answer"] = r["answer"].replace(code_recommendations, "")
```

When the recommendations contained a ```json fence, `code_recommendations` was rebound to the unwrapped inner content. The `.replace()` then matched only that inner content as a substring inside the original fence, leaving the wrapping `<CODE_RECOMMENDATIONS>`, the ```json fence itself, and the closing tag in `answer`.

Fix: capture the original (pre-unwrap) inner text, splice the full `<CODE_RECOMMENDATIONS>...</CODE_RECOMMENDATIONS>` block out of the answer, then unwrap the inner fence for the structured `code_recommendations` field.

3. `_assess_factuality_issue` raised when the LLM omitted recommendations

The prompt explicitly says "If you have recommendations to fix or update the code, add a JSON object inside an xml tag…" — recommendations are optional. But `extract_xml_tag` was called unconditionally, so a successful assessment with no recommendations surfaced as a tool error to the MCP host. Now caught and returns `code_recommendations: ""`.

4. Prompt typos + invalid example JSON

Both prompts had typos ("asessment", "udpate", "attemps", "You task"). The assess prompt also showed the model an example schema using single-quoted "JSON" with no commas between objects — invalid JSON. LLMs that follow the example faithfully produced strings the downstream JSON parser would reject. Rewritten to be valid double-quoted JSON. Similar fixes to the extract prompt (implicit string concatenation, missing array commas).

Tests

• `TestExtractTag` — added `test_does_not_span_multiple_fences` (regression for Submoduling #1).
• `TestAssessFactualityIssue` — rewrote assertions: the existing `test_strips_recommendations_block_from_answer` was pinning the broken behavior (`assert "<CODE_RECOMMENDATIONS>" in result["answer"]`), so it's been flipped to assert the cleaned shape. Added `test_returns_empty_recommendations_when_tag_absent` (cleanup: remove EVEMCPClient and expand README for v0.0.1 #3) and `test_keeps_malformed_inner_fence_as_is` (defensive).

Coverage stays at 100%.

Test plan

• `poetry run pre-commit run --all-files`
• CI matrix
• Manual smoke: `assess_factuality_issue` with a real GEE script after v0.0.1 lands

[r-spiewak]

I didn't test it, but there are some minor additions to your tests.