fix(asap-tools): score SQL query results in calculate_fidelity

[milindsrivastava1997]

Closes #885.

Problem: SQL results are stored as TabSeparated text in raw_text_result, and get_all_timeseries only read result, so calculate_fidelity never scored a SQL query.

Fix:

• parse_tabseparated_result(query, raw_text) turns TSV rows into the same {frozenset(label → value): float} shape result uses.
• Label names come from the query's outermost SELECT list. Bare columns are labels, and the single other item is the value.
• QueryResult.keyed_result() returns result, or the parsed TSV; get_all_timeseries uses it. Parsing happens at scoring time, so loaded JSONL results are covered too.
• SELECT-list names (not positions) mean a ClickHouse TSV exact result also matches an estimate in Prometheus JSON, e.g. from the rqe route (feat(asap-tools): rqe query protocol and frontend switch for e2e runs #890).
• Malformed results raise instead of being skipped: wrong column count, duplicate label set, non-numeric value, or not exactly one aggregate.

Known limitation: ASAPQuery's ClickHouse adapter writes label columns in alphabetical order (KeyByLabelNames::new sorts them), while ClickHouse uses SELECT order. If a query's GROUP BY columns aren't alphabetical in its SELECT list, an ASAPQuery TSV estimate's series won't match the exact series and get skipped; in rare cases values could pair with the wrong labels. Answers through the rqe route are unaffected, since they carry label names. A follow-up could make the adapter emit SELECT order.

Also not covered:

• the parquet serializer drops raw_text_result;
• recommended_sketch_configs.py and kll_rank_error.py still read .result;
• ClickHouse's escaping of tabs and backslashes inside values isn't undone.

Tests: asap-tools/experiments/tests/test_calculate_fidelity.py (13).

🤖 Generated with Claude Code