evaluate --exit-code: escape quotes (or emit JSON Lines) in LLM-judge FAIL feedback field

[caohy1988]

Context

PR #45 (merged) verified the LLM-judge AI.GENERATE path end-to-end against live BigQuery. During that smoke, one FAIL line surfaced an unescaped embedded double-quote inside the feedback="..." value when the judge's justification itself contained quoted text:

FAIL session=4ca31e85 metric=faithfulness score=0.6 feedback="The agent added "(Design)" to Jordan Lee's name, which was not present in the user's request or any provided context."

That parses fine for a human eyeball but breaks awk -F'"', cut -d'"', or any shell parser that splits on " to extract the feedback field. Apostrophes inside the feedback (Jordan Lee's) also wouldn't survive a switch to single-quote wrapping.

Why this is not a blocker

• Doesn't affect SDK correctness, AI.GENERATE behavior, or judge scoring.
• Doesn't affect blog post Revise README for clarity and updated link #3 (Docs: Medium blog post #3 plan — LLM-as-Judge in BigQuery (some regressions don't fit a budget) #82) — the human-readable CI screenshot looks correct, the cover demo's value is unchanged.
• Pass/fail outcome and exit code are unchanged.

Tracking as a follow-up so it doesn't get lost.

Two options

1. Escape " and \ inside the feedback value. Single-line, minimal change to cli._format_feedback_snippet (or to the FAIL-line composition in _emit_evaluate_failures):

   escaped = feedback.replace("\\", "\\\\").replace('"', '\\"')
   parts.append(f'feedback="{escaped}"')

   Pros: smallest delta; preserves the current key=value shape readers may already grep.
   Cons: still a hand-rolled DSL — anyone parsing it has to know about the SDK's escape convention.

2. Add an opt-in JSON Lines mode for CI parsing. New flag like --exit-code --emit=jsonl that emits one JSON object per failing (session, metric) pair to stderr instead of the current key=value line. Default stays human-readable.

   Pros: CI integrations get a proper machine-parseable contract; no escape ambiguity ever; future fields (e.g. multi-metric judge results, criterion arrays) can grow without breaking parsers.
   Cons: bigger surface; needs a flag-shape decision; existing readers stay on the human format anyway.

Recommendation

Ship Option 1 first (~10-line change, drops in next polish window). Track Option 2 as a separate larger ask if a CI consumer actually needs the JSONL contract — don't pre-add the surface for a hypothetical use.

Surface to touch

• src/bigquery_agent_analytics/cli.py — _format_feedback_snippet (or the line composition in _emit_evaluate_failures).
• tests/test_cli.py::TestFormatFeedbackSnippet — add an escape-roundtrip case.
• tests/test_cli.py::test_evaluate_exit_code_llm_judge_emits_feedback_snippet — extend the assertion to cover an embedded-quote justification (use the real (Design) example from the PR chore: Configure Renovate #45 smoke as the regression seed).

Ref

• PR chore: Configure Renovate #45 (live AI.GENERATE rewrite): fix(llm-judge): make AI.GENERATE template execute against current BigQuery haiyuan-eng-google/BigQuery-Agent-Analytics-SDK-fork#45
• Blog post Revise README for clarity and updated link #3 plan: Docs: Medium blog post #3 plan — LLM-as-Judge in BigQuery (some regressions don't fit a budget) #82
• Series plan: Docs: Medium blog post series plan — downstream value for BQ Agent Analytics users #51