Profine

Telemetry bugfix release. Two NULL columns the priors flywheel needed were never being emitted; the materialized view's success_rate column has been NULL for every user since telemetry shipped, silently breaking the suggester's failure-avoidance filter.

pip install -U profine

Fixed

• _loss_ok_from_bench was reading a non-existent key. The function read correctness.verdict from benchmark_comparison.json, but the correctness sub-dict has a passed (bool) key, not a verdict. (verdict is a top-level BenchmarkComparison field.) The bug emitted loss_ok=None for every run sent to the telemetry backend. The test fixture was buggy in the same way as the production code (it also used the verdict key), explaining the silent regression. Adds a focused regression test (test_loss_ok_reads_correctness_passed_not_verdict) that pins both the real and old-buggy shapes.
• _gather_outcomes never set runtime_seconds. The field is in ALLOWED_OUTCOME_FIELDS but was missing from every emit-row dict literal, so run_optimizations.runtime_seconds was NULL across the entire dataset. The primary outcome now carries the profile's wall-clock runtime; stacked/skipped rows still omit it (no per-entry attribution available — same convention as speedup_factor).