v3.11.0

Added

• orq-analyze-traces/scripts/decide.py (RES-1629): one POST /v3/router/decisions call per trace over every trace in the window, to surface failing traces the analyst never read. Traces go whole to typesafe/jev-latest when they fit and to openai/gpt-6-luna otherwise, never chunked or trimmed. Subcommands probe, estimate, label (resumable, capped by a required --budget-usd, refuses windows over 5,000 traces without --max-traces), pick, agreement and rates. Never writes to the platform.
• orq-analyze-traces: a Phase 1 selection pass over the window, a Phase 4 calibration against the analyst's own labels with a fixed clearing rule (10+ agent positives, recall 0.7+), and a window pass that reports each cleared mode's flag rate, interval and flagged trace ids. Spend is approved through AskUserQuestion. Falls back to the sample-based path when decisions are unavailable or declined.

Changed

• Error-analysis artifact: rate is now the best available rate and rate_basis says which (window_flags or read); read_rate always holds the read-based rate. New decisions, traces.labelled, per-mode window_flags and flagged.
• orq-improve-agent, orq-recommend-evaluators: quote rate with its rate_basis, and treat flagged as unread examples, never as regression evidence.

Fixed

• Calibration counts missing analyst-positive labels as misses and rejects unreadable probabilities. Budget reservations halt the whole pass on refusal, retain uncertain charges and attempted overflow retries, and reject non-finite limits. Active-window metadata is written before labelling; reports and picks filter the cache to that population and refuse missing metadata. Add an offline regression suite and pinned API contract sources.
• Preserve complete API-returned trace IDs in analysis evidence, including platform ULIDs as well as OpenTelemetry IDs; do not substitute record or span IDs.