Sub-Subnets

[gztensor]

Subsubnet Support: Multi-Weight Spaces per Subnet

This PR introduces subsubnets, a new abstraction layer within subnets that enables multiple independent weight spaces, incentive flows, and emission configurations within the same subnet. Each subnet can define up to 8 subsubnets (IDs 0–7), allowing complex multi-task validation while maintaining validator cohesion.

Key Features:

• Per-subsubnet weight setting, commitment, and emission logic
• Fully backward-compatible with legacy miners and validators (via subsubnet_id = 0)
• Automatic cleanup of weights when subsubnet limits are reduced
• Configurable emission proportions (default: Fibonacci distribution)
• Independent calculation of trust, consensus, and incentives per subsubnet
• Strict write permissions enforced for subnet owners and sudo only

Governance Capabilities:

• Subnet owners can configure desired subsubnet limits and emission proportions
• Global subsubnet limit caps usage across all subnets and triggers cleanup when reduced
• Validator-majority (kappa) support for voting on incentive distribution logic

Purpose:
This functionality enables richer subnet behavior, supports multi-objective tasks, and introduces a scalable framework for task-specific mining without increasing subnet count.

Refer to the linked BIT and test plan for detailed specification and test coverage.

[ppolewicz]

both the speed of increase and the speed of decrease should be limited

[ppolewicz]

The emission should by default use reverse fibonnaci, not fibonnaci, so that the subsubnet with the highest emission has the lowest id, so that if the governor needs to decrease the global limit in an emergency (however unlikely it is), the highest emission subsubnets prevail. In a future version when we'll have configurable subsubnet emission proportion, the emission proportion should be sorted in decreasing order for the same reason.

[ppolewicz]

I propose the limit imposed by the governor for a given subnet to be calculated as max(1, math.floor(global_subsubnet_per_subnet_limit * (256/max_uids)), so that for sn1 it'll shrink by a factor of 4

[ppolewicz]

In v1 the number of knowledge commitments should not scale with subsubnets, but in a future version it totally should.
The knowledge commitment system already supports multiple slots, so far we've just used slot zero. When effective subsubnet limit decreases, the knowledge commitments with ids higher than the maximum currently supported should be wiped out.

[ppolewicz]

A possible future extension, idea of @adriansmares: allow miners to set separate neurons for each subsubnet

[ppolewicz]

For serious consideration of including in scope of the first version, an idea by Fish is to make the pruning score proportional to the subsubnet emission, so that the miner second-in-terms-of-emission on the lowest emission subsubnet may be (if he's close enough to the best miner in that subsubnet) deregistered after a second-in-terms-of-emission miner in the highest emission subsubnet, if that one is not close enough to the top one in his subsubnet.