feat(storage-bench): storage/DB backend perf + integrity benchmark, aligned to the storage standard

[mdheller]

A benchmark suite for storage / DB backends where integrity is a first-class result, aligned to SocioProphet/socioprophet-standards-storage (evidence-first: a number is a claim only if backed by a reproducible record).

• ts/src/storage-bench.ts — benchObjectBackend measures any ObjectBackend (InMemory / S3-BYOS / RocksDB / …) on put / get / verify: latency p50/p95/p99 + throughput (ops/s, MB/s), over deterministic seeded content-addressed blobs. Fail-closed: a blob that doesn't byte-round-trip, or whose content hash doesn't re-derive, fails the run — a fast backend that returns the wrong bytes fails here, by design. toEvaluationRecord emits a record conforming to evaluation-record-standard.v1 (subject_type=platform_capability, attempt_mode=benchmark_run), carrying the raw measurements as evidence.
• storage-bench.test.ts — 5 tests, teeth both ways: correct backend passes with full integrity + ordered percentiles; deterministic for a fixed seed; a corrupting backend fails; a losing backend fails; the EvaluationRecord conforms.
• docs/STORAGE_BACKEND_MATRIX.md — the Table-2.1 analog: estate backends (Canonical object store, S3-BYOS, RocksDB, Hypercore, AtomSpace, FederatedAtomSpace) + the external field, scored on the standard's dimensions (content-addressed / provenance / integrity-at-rest·codex / fault tolerance / sovereign-BYOS / distributed / fail-closed verify). Perf numbers live in EvaluationRecords, never inline.

Typecheck clean; suite 5/5; dist rebuilt. Runs under the existing agent-machine/ts-ci test step.

🤖 Generated with Claude Code

[mdheller]

Merge rationale (self-hosted CI queued, not failing): ran the full CI-equivalent locally — npm ci && npm run build (dist in sync, the build-and-verify-dist assertion) + full ts suite 471/471 pass + tsc --noEmit clean. Change is additive (new storage-bench module + test + index export + rebuilt dist + a doc); it touches no existing module.

Adversarial self-review: integrity is fail-closed and proven both ways — a corrupting backend (wrong bytes) and a losing backend (missing blob) both FAIL the run, a correct backend passes with full integrity; runs are deterministic for a fixed seed; the EvaluationRecord conforms to evaluation-record-standard.v1 (evidence carried, no bare numbers). Merging on the local proof.