docs: fix evaluator variable mappings that silently break judge scores

[tkykenmt]

Fixes #30

Summary

Fixes three evaluator variable mappings in the workshop docs that break silently — the judge runs with an empty (or wrong) variable and still returns a plausible-looking score.

• 06 Experiments (Correctness, learner + instructor): restore the slice form $.messages[-1:].content for {{query}}. Clarify experiment evaluator variable mapping #15 rewrote the table from the slice form to a bare $.messages[-1].content while renaming the columns, and the evaluator's JsonPath engine (jsonpath-plus, eval: false) silently resolves bare negative indexes to nothing. Also adds warnings about leading whitespace in JsonPaths and a tip to debug via the langfuse-llm-as-a-judge execution traces.
• 04 Monitoring (Out-of-Scope Request, learner + instructor): map {{last_user_message}} to $.messages[1].content. The evaluator targets the final generation of the tool-calling loop, whose input ends with tool results, so the [-1:] slice handed the judge the tool payload and obvious out-of-scope requests were scored as in-scope. The instructor notes already listed this pitfall; both docs now carry the fix.
• 06 Experiments (keyword_overlap): note that the out-of-scope items' expectedKeywords test literal wording, so a good refusal phrased differently can score 0 on them.

Verification

Confirmed two ways:

1. On Langfuse Cloud, by inspecting the compiled judge prompts in the langfuse-llm-as-a-judge execution traces before/after each mapping change (empty Query: / Ground truth: before; filled and scoring correctly after).
2. Standalone, with jsonpath-plus@10.3.0 invoked the same way as packages/shared/src/features/evals/utilities.ts.

[CLAassistant]

All committers have signed the CLA.