feat(request-trace-uploader): add configurable per-record-type sampling

Why

High-volume deployments cannot export every record. Sampling must be coherent: a request's payload record and its metadata record have to be kept or dropped together, or the export is useless for both audit and replay.

Scope

• Apply sampling after record suppression. Suppressed and non-suppressed records are sampled identically.
• Declare the sampling key per record type rather than as a cross-type fallback chain. request_end uses request.request_id, request_payload uses payload.request_id, and the three tool events use agent_context.session_id. A fallback chain across these would mix identifier namespaces and break coherence, because a session contains many requests.
• Make the decision a deterministic function of the key alone, so records sharing a key are kept or dropped identically across record types, segments, and uploader restarts. Per-record randomness is not acceptable: a request's payload and its request_end are written to the same file family but may land in different segments uploaded at different times, and determinism is what makes them agree without shared state.
• Document that tool events carry no request identifier, so tool sampling is coherent at session granularity and does not join to request-level sampling.
• Define policy for records with no resolvable key: keep, drop, or reject at config validation. Default keep.
• Emit records_sampled_out_total by record type and reason, where reason is rate or no_key.

Tests

• Deterministic keep or drop for the same key across record types, segments, and process restarts.
• A request's request_payload and request_end receive the same decision.
• Tool events key on session and are unaffected by request-level decisions.
• Missing-key policy under each setting.
• Rate boundaries at 0.0 and 1.0.

Note

Record-level sampling requires rewriting each segment, which forfeits byte-exact jsonl.gz passthrough on the object-store backend. If passthrough matters more than granularity, segment-level sampling is the alternative, but it cannot give coherent per-request retention.

Parent: #1004

The issue names no files or existing tests; start by locating the request-trace uploader and definitions for request_end, request_payload, and the tool events. Read parent issue #1004, then identify where suppression, segment rewriting, configuration validation, and object-store passthrough are handled. Done means deterministic per-key sampling, missing-key policy, metrics, documentation, and the listed boundary and cross-record tests.