test(promql-compliance): add olly-bench suite and p75 quantile queries

[milindsrivastava1997]

Summary

• Add suites/olly-bench.yaml + datasets/olly-bench.yaml: the 25 unique olly-bench handwritten_v2 queries tracked in Tracking: asap-planner query coverage for olly-bench #738. Query structure is unchanged; durations are shrunk to fit the fixture ([1h]/[6h]/[12h]/[24h] → [5m], offsets → 5m, [6h:] → [10m:]). Duplicate lines 17 and 22 are omitted; names carry the original line numbers.
• Split the runner Makefile: make run-all runs every suite, make run-for-ci runs only CI_CASES. CI now uses run-for-ci, since most olly-bench queries are expected to fail until Tracking: asap-planner query coverage for olly-bench #738 lands.
• Add p75 quantile_over_time / quantile / quantile by (job) / quantile by (job, instance) queries to the quantiles suite.
• Register the olly-bench pair in the checked-in fixture parse test.

Notes

• Dataset is generated by a Python script (not checked in); the formulas are documented in the YAML header. No series or jobs tie, so topk/sort_desc have unique answers, and job worker has zero 5xx so > 0 actually filters.
• sort_desc ordering is not verified: the comparator sorts series before comparing.

Test plan

• go test ./... in runner and seeder
• Live quantiles suite: passes, including all four p75 queries
• Live olly-bench suite: 3/25 pass (lines 5, 6, 9), matching Tracking: asap-planner query coverage for olly-bench #738; the rest fail with No result for query from ASAPQuery. Prometheus reference parity passes on every query.

Refs #738

🤖 Generated with Claude Code