…backoff

- Detect 'restart required' and 'connection closed' errors for immediate reconnect
- Change base reconnect delay from 30s to 10s with exponential backoff (10s->20s->40s->80s->120s max)
- Reset backoff counter on successful reconnect or receiving blocks
- Add reconnect_failures counter for tracking consecutive failures

When Bittensor RPC connection errors occur (e.g., 'restart required'),
the reconnection logic now properly stops the old BlockSync before
creating a new one. This prevents the old internal task from
continuing to emit errors.

Key changes:
- Track BlockSync instance instead of just the spawned task handle
- Call sync.stop() before creating new connection to clean up internal tasks
- Remove unnecessary wrapper task around sync.start()
- Maintain exponential backoff for reconnection attempts

This resolves the issue where connection errors would spam logs every
5 seconds due to the old BlockSync task not being properly terminated.

…sor-reconnection-cleanup

fix(bittensor): improve reconnection by properly stopping old BlockSync

…ases sled lock

The test was flaky due to potential race condition in nextest parallel execution.
Added explicit flush() call and drop() to ensure sled database is properly
flushed and the lock is released before reopening.

…ormNetwork#34)

Add proportional normalization for challenge weights to prevent
any single challenge from exceeding its allocated weight share.
When a challenge returns weights that sum to more than 1.0,
each weight is scaled down proportionally to sum exactly to 1.0.

- Add normalize_hotkey_weights() function
- Integrate normalization in collect_challenge_weights()
- Add comprehensive unit tests for edge cases

…ormNetwork#35)

Add term-llm-reviewer: prefix to the allowed image tags list alongside
term-compiler: and ghcr.io/ prefixes. This fixes the policy violation
error when building LLM reviewer images for agent evaluation.

…(UID + hotkey + stake)

The metagraph handler (api::metagraph::get_metagraph) existed but was never
registered in the router, causing 404 errors when term-challenge tried to
refresh its neuron cache from the platform server.

Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>