Fix/per block runtime decoding

[nol4lej]

Summary

Release 0.7.1, with no API changes.

• Fix: each block is decoded with its own runtime's metadata. getDynamicBuilder(blockHash?) and getExtrinsicDecoder(blockHash?) fetch the metadata of the runtime a block executed under, over RPC, once per spec_version. queryBlockEvents passes its hash. Before, the metadata cached at connect decoded every block, so at spec 17 every earlier shield misdecoded, and a long-lived session misdecoded blocks after an upgrade. Without a hash, behaviour is unchanged.
• Fix: the shield commitment is validated on every path. assertShieldParams requires a 32-byte hex commitment. The pallet call now refuses what the precompile calldata already refused.
• Refactor: SubstrateClient is split by responsibility.
  • runtimeDecoders: metadata and decoders per runtime. Each decoder is built on first use, and a failed fetch is retried instead of cached.
  • eventRecords: System.Events → EventRecord.
  • blockInfo: block time from timestamp.set and author from the digest.
  • getBlock is split into resolveBlockHash and readTimestamp.
  • The connected runtime's metadata is now fetched once for both decoders, instead of once by each.
  • The public API is unchanged: extractAuthorFromLogs and the DynamicBuilder / ExtrinsicDecoder types are still exported from SubstrateClient.

Test plan

• pnpm run check: 899 passed
  • per-runtime cache
  • concurrent first fetch downloads metadata once
  • retry after a failed fetch
  • unknown block is not cached
  • only the requested decoder is built
  • timestamp fallback edge cases
  • commitment validation
• pnpm audit: no known vulnerabilities
• pnpm pack --dry-run
• Against a spec 17 dev node: every extrinsic and the events of every block decode with the block hash. The indexer and privacy-explorer read shields correctly.