fix(results): isolate target bundles under run timestamps

[christso]

Summary

Multi-target eval runs now fan out into isolated target/variant result bundles under the invocation timestamp, so two targets with the same test_id no longer point at and overwrite the same sidecar files. Within each bundle, row sidecar directories use deterministic compact row IDs based on eval source, suite, test id, target, and variant, while index.jsonl remains the source of truth for identity and artifact paths.

Dashboard and result readers now discover nested bundle manifests under timestamp folders while preserving legacy root-level index.jsonl bundles. They may walk storage folders to find manifests, but target and variant semantics continue to come from loaded row/run metadata, not path parsing.

Fixes #1557.

Validation

• Rebased onto current origin/main; branch freshness check git log HEAD..origin/main is empty.
• bun install after rebase to materialize declared dependencies.
• bun run build
• bun run typecheck
• Focused tests for eval artifact writing, aggregation, nested Dashboard discovery/drilldown, export, rerun, integration output paths, and core evaluation APIs: 394 passed, 0 failed.
• Earlier full local bun run test: core 2063 passed, CLI 723 passed, SDK 88 passed, Dashboard 142 passed.
• Local two-target mock eval dogfood: produced distinct mock-alpha/index.jsonl and mock-beta/index.jsonl bundles with separate row IDs and answer files.
• Same-test_id/duplicate-suite dogfood: produced two distinct row sidecar directories inside one target bundle and preserved both summary rows.
• Dashboard browser/API UAT against a canonical nested-bundle project: list, detail, file drilldown, and compare all loaded both targets from metadata.
• Remote sync dogfood with a file-backed agentv/results/v1 branch: pushed nested bundles, cleared local materialized results/cache, synced through Dashboard project APIs, and loaded remote list/detail/compare/drilldown successfully.
• Live provider/grader dogfood through the local OpenAI-compatible OAuth proxy: codex target with api_format: responses plus openai grader target with api_format: chat, all credentials/model routed through LOCAL_OPENAI_PROXY_* env refs. Result: PASS, 1/1, score 1.0. Run bundle: /tmp/agentv-av9vi-live-utlpOx/.agentv/results/av9vi-live-dogfood/2026-06-29T02-51-25-032Z/codex-local-proxy/index.jsonl; row sidecar: live-proxy-case--3d1e8b8bde59/run-1/; grader type: llm-grader; answer: agentv live dogfood ok.

---

Private evidence: https://github.com/EntityProcess/agentv-private/tree/evidence/av-9vi-result-row-sidecars (commit 0b585d2) preserves the live local-proxy Codex target + OpenAI grader run bundle on an orphan branch.

[cloudflare-workers-and-pages]

Deploying agentv with    Cloudflare Pages

Latest commit:	f656fd4
Status:	✅  Deploy successful!
Preview URL:	https://e72a68d4.agentv.pages.dev
Branch Preview URL:	https://result-row-id-sidecars.agentv.pages.dev

View logs