feat: add JUnit XML exporter and --fail-under CI quality gate

[brainsparker]

What this does

Adds CI-native quality reporting to PromptLens: a JUnit XML exporter and a --fail-under score gate, so eval runs can block a pipeline the same way unit tests do.

New output format:

output:
  formats: [html, json, junit]

New CLI gate:

promptlens run config.yaml --fail-under 3.5

Behavior:

• One JUnit test suite per model, one test case per golden-set entry.
• A test case whose judge score is below the threshold is a failure, a model API error is an error, and an unjudged case is skipped (so judging-disabled runs never report a false pass).
• Suite properties carry run metadata CI dashboards can surface: model, provider, run id, golden set, average judge score, total cost.
• If any model's average judge score falls below --fail-under, the run exits with code 2. Exit code 1 stays reserved for run errors, so CI can distinguish quality regressions from infrastructure failures.
• promptlens export <run_id> --format junit re-exports any saved run.

Why now (market rationale)

1. OpenAI acquired Promptfoo (announced March 9, 2026) and is folding it into its enterprise agent platform with a security and red-teaming focus (https://openai.com/index/openai-to-acquire-promptfoo/, https://techcrunch.com/2026/03/09/openai-acquires-promptfoo-to-secure-its-ai-agents/). Community coverage flags real anxiety about a formerly vendor-neutral OSS eval tool now owned by a model vendor. That is a genuine opening for lightweight, vendor-neutral eval tools, which is exactly PromptLens's positioning.
2. JUnit XML output is now table stakes across the category. Promptfoo shipped and documented it as a flagship output format in 2026 after it sat as a long-standing feature request (https://www.promptfoo.dev/docs/configuration/outputs/, JUnit-style XML reports promptfoo/promptfoo#9071). DeepEval is winning the CI/CD lane largely on pytest-native and JUnit-compatible reporting, and newer tools (tribunal, ask-eval, evalci) all treat JUnit plus exit-code gating as baseline.
3. The gate is what makes evals actionable: baseline-and-block patterns (fail the build when scores drop) are the recurring ask across 2026 eval tooling (see multivon-eval's ci_eval example and llm-eval-suite's regression flow). --fail-under is the minimal version of that, and it works with every existing format, not just junit.

This lane was chosen deliberately to avoid the repo's open-PR congestion: the change is a new exporter module plus small additive wiring in cli.py. It does not touch promptlens/models/config.py or the provider HTTP code, which is where the 30+ open nightly hardening PRs overlap. The formats list in OutputConfig is an unvalidated List[str], so no config model changes were needed.

Runner-up candidates considered today and passed on: a spec-compliance feature for MCP-Profiles around the new MCP list-result caching hints (SEP-2549 ttlMs/cacheScope, https://blog.modelcontextprotocol.io/posts/2026-07-28/), dropped because that repo has pivoted to the you-aware search server where gateway caching semantics no longer fit; and a regression baseline comparison command for PromptLens (promptlens compare --baseline), a strong follow-on that builds naturally on this PR's exit-code convention.

Implementation notes

• promptlens/exporters/junit_exporter.py: new JUnitXMLExporter following the existing BaseExporter interface, built with xml.etree.ElementTree (proper escaping, no string templating). Configurable fail_under threshold, default 3.0 on the 1-5 judge scale.
• promptlens/exporters/init.py: registers the new exporter.
• promptlens/cli.py: junit entry in both the run formats map and the export command, new --fail-under option (click.FloatRange(1.0, 5.0)), and a _check_fail_under helper that reports every failing model with its average score. A model with no judge scores fails the gate explicitly rather than passing silently.
• tests/test_junit_exporter.py: 18 new tests covering pass/failure/error/skipped mapping, custom thresholds, multi-model suites, XML escaping, suite properties, directory creation, and the gate logic (including the exactly-at-threshold and no-scores cases).
• README.md: feature list, output config, and a new CI/CD Integration section with a GitHub Actions example.

Test status

• Full suite: 27 passed (18 new, 9 existing), Python 3.9, pytest.
• All touched modules compile clean and the CLI help renders correctly.
• End-to-end verification: exported a saved run via promptlens export --format junit and validated the XML parses, with correct failure counts per suite.

[brainsparker]

Closing in favor of #43, the same change set rebased onto post-consolidation main (the cli.py conflict with #40 is resolved there). Full suite passes on the rebase: 54/54.