test: add large PR review harness

[rianjs]

Closes #375.

Summary

• add scripts/verify-large-pr-review.sh for large-PR no-post review regression checks
• add a build-tagged JSON field helper for structured cr review --json extraction
• validate artifact shape, source sentinel absence from model-facing artifacts, prompt/context artifact size limits, stderr LLM task breadcrumbs, and full durable-task reuse across two unchanged passes
• include --self-test for credential-free positive and negative invariant checks

Verification

• rtk scripts/verify-large-pr-review.sh --help
• rtk scripts/verify-large-pr-review.sh --self-test
• rtk go test ./internal/pipeline ./internal/cmd/reviewcmd ./internal/cmd/benchmarkcmd -count=1
• rtk make lint
• rtk make build

Empirical Harness Note

The real harness mode is dry-run/no-post and isolates data/cache state while using the maintainer's normal cr config. I started one local real pass with CR_LARGE_PR_KEEP_TMP=1; it blocked during local runtime construction before PR fetch or LLM breadcrumbs, so I terminated that harness process tree rather than treating it as a product signal. The script now has a per-pass timeout to make that failure mode explicit.