feat(trace): add normalized trajectory contract

[christso]

Summary

Trace importers and future trajectory-aware graders now have one versioned contract to share. The new normalized trajectory model preserves ordered events, tool identity, branch selection, timing provenance, redaction state, raw evidence handles, and source references while still projecting back to the existing compact TraceSummary shape.

This stays inside the foundational model bead: no replay target database, no replay CLI alias, and no OTLP/Phoenix/Pi importer wiring is introduced here.

Design Notes

• Persisted trajectories use schema_version: agentv.trace.v1 and snake_case wire keys; TypeScript internals remain camelCase.
• Branchable sources can carry explicit includedEventIds, so summary derivation only grades the selected path when one is provided.
• Tool status/error preservation is explicit: error, timeout, and cancelled count as summary errors, while unknown is retained without being treated as a failure.
• @agentv/eval now exports matching camelCase Zod schemas for downstream SDK authors without changing code-grader input plumbing in this bead.

Verification

• bun test packages/core/test/evaluation/trace-trajectory.test.ts packages/core/test/evaluation/trace-summary.test.ts (23 pass)
• bun node_modules/.bin/biome check packages/core/src/evaluation/trace.ts packages/core/test/evaluation/trace-trajectory.test.ts packages/eval/src/schemas.ts packages/eval/src/index.ts
• bun --filter @agentv/core typecheck
• bun --filter @agentv/eval typecheck
• bun run typecheck
• bun --filter @agentv/eval test (67 pass)
• bun --filter @agentv/core test (1828 pass)
• bun --filter @agentv/core build
• bun --filter @agentv/eval build

---

[cloudflare-workers-and-pages]

Deploying agentv with    Cloudflare Pages

Latest commit:	553b61e
Status:	✅  Deploy successful!
Preview URL:	https://3ce0aee0.agentv.pages.dev
Branch Preview URL:	https://feat-av-vwa-4-normalized-tra.agentv.pages.dev

View logs

[christso]

Review/rework for the TraceSummary vs NormalizedTrajectory concern is complete.

Decision: keep both, but as one canonical model plus one derived read model. NormalizedTrajectory is the full source of truth for trajectory artifacts, replay, and future trajectory-aware grading. TraceSummary remains the backward-compatible compact result/grader/dashboard read model and must be derived from a full trajectory when one exists.

Primary-source rationale:

• Terminal-Bench keeps result models and recordings/logs separate: TrialResults/BenchmarkResults carry result fields plus recording_path, the harness writes run-level results.json, writes sessions/agent.cast separately, and documents individual trial results.json files as source of truth while the run-level results.json is rebuilt aggregation. See https://github.com/laude-institute/terminal-bench/blob/1a6ffa9674b571da0ed040c470cb40c4d85f9b9b/terminal_bench/harness/models.py#L43-L64, https://github.com/laude-institute/terminal-bench/blob/1a6ffa9674b571da0ed040c470cb40c4d85f9b9b/terminal_bench/harness/harness.py#L784-L831, and https://github.com/laude-institute/terminal-bench/blob/1a6ffa9674b571da0ed040c470cb40c4d85f9b9b/terminal_bench/harness/harness.py#L1039-L1090.
• Harbor, the Terminal-Bench successor framework, shows result.json beside agent/recording.cast and agent/trajectory.json, and the viewer separately inspects trial results, trajectories, metrics, artifacts, and summaries: https://www.harborframework.com/docs/run-jobs/run-evals. Its SFT docs require ATIF trajectory format while success filtering uses result.json: https://www.harborframework.com/docs/training-workflows/sft.
• OpenInference and OpenTelemetry model full execution as traces/spans with LLM/tool/token attributes, while observability tools expose derived views, scores, metrics, and raw trace data over that trace substrate: https://arize-ai.github.io/openinference/spec/, https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-spans/, https://www.braintrust.dev/docs/observe/examine-traces, https://langfuse.com/docs/observability/data-model, https://arize.com/docs/phoenix/tracing/llm-traces.
• OpenAI Evals API returns run-level result counts and granular testing criteria results with a report URL, which is consistent with a result/read-model surface over execution evidence: https://platform.openai.com/docs/guides/evals.

Changes pushed in 553b61e60896ab9971a0f175c848261f1233774f:

• Reworded docs/plans/trace-evaluation-architecture.md from ambiguous two-layer wording to one canonical trajectory plus derived read models.
• Updated packages/core/src/evaluation/trace.ts and packages/eval/src/schemas.ts comments so TraceSummary is explicitly derived compatibility/read-model state.
• Added a focused regression test that normalized trajectory wire state does not embed trace, summary, or trace_summary, then derives the expected TraceSummary from the full trajectory.
• Added the same source-backed decision and verification to Bead av-vwa.4.

Verification:

• bun test packages/core/test/evaluation/trace-trajectory.test.ts -> 9 pass
• bun node_modules/.bin/biome check docs/plans/trace-evaluation-architecture.md packages/core/src/evaluation/trace.ts packages/core/test/evaluation/trace-trajectory.test.ts packages/eval/src/schemas.ts -> no fixes
• bun --filter @agentv/core typecheck -> pass
• bun --filter @agentv/eval typecheck -> pass
• git diff --check -> pass
• ce-code-review pass over the focused diff found no blocking issues

Residual risk: later importer/replay/grader beads must preserve this invariant when they add persistence or wiring. PR #1331 remains unmerged.