fix(ai): simplify trace validation and align request status

[ice-black-tea]

Summary

• fix RuntimeHistory.trace() rejecting valid tool events that carry request_sequence without request_purpose
• remove redundant projection-layer validation while retaining checks required to decode, correlate, or disambiguate persisted data
• keep trace usage metadata tolerant of missing/additive fields while rejecting malformed present numeric values
• persist model-request cancellation as model_request_cancelled and expose trace status as CANCELLED, matching ModelInteraction/metrics semantics
• normalize already-persisted cancellation events from the previous writer without a data migration
• add regression coverage for real tool-turn trace replay, sparse lineage metadata, usage projection, and cancellation status

Validation ownership

• writers own completeness of model/tool observability facts
• history/trace readers validate only values they must parse and identities they must correlate
• opaque trace strings such as purpose and observation id are forwarded rather than restricted by the reader
• public ExecutionTraceItem is a transport DTO and no longer duplicates model-specific payload validation
• cursor bounds, run/segment coordinates, duplicate lifecycle events, malformed numeric metadata, and ambiguous tool correlation remain fail-closed

Compatibility

• no database schema or migration changes
• no package or MCP version changes
• no public API removals
• existing sequence-only tool events become readable without migration
• existing cancellation records written as model_request_failed are normalized to CANCELLED when they carry the previous writer's cancellation signature

Validation

• static review completed against master@0765c40546f4c870055cd063435e0cdae89357c4
• targeted regression coverage added in the existing test suite
• latest PR CI passed on 3dc5db727ed846449538702e156c0f53e051bb46