feat: more paranoid reth db checks

[dvush]

Now we check reth db account trie table for correct state root of the parent instead of checking latest block.

📝 Summary

💡 Motivation and Context

---

✅ I have completed the following steps:

• Run make lint
• Run make test
• Added tests (if applicable)

[claude]

Claude finished @dvush's task in 1m 14s —— View job

---

Code Review: feat: more paranoid reth db checks

• Read and analyze changed files
• Review code quality and correctness
• Post review feedback

---

Summary

This PR replaces the block-hash-based DB consistency check in v2 with a state-root-based check. Instead of verifying last_block_hash against the latest block in the DB, it now computes the state root via StateRoot::from_tx().root() and compares it to the expected parent_state_root.

Issues

1. Performance — state root recomputation is expensive (inline comment posted): check_state_root_in_db calls StateRoot::from_tx(provider.tx_ref()).root(), which walks through trie cursors to compute the state root. This is called once per storage proof target (in into_par_iter) and once for account proofs. The old check was a cheap block hash/number lookup. This could be a meaningful regression for blocks touching many accounts.

2. v_experimental still uses the old last_block_hash pattern: v_experimental/fetch.rs and v_experimental/mod.rs still use last_block_hash with the old block-hash-based check. Was this intentional (only upgrading v2), or should v_experimental be updated to match?

3. Minor inconsistency — missing semicolon at line 119: check_state_root_in_db(&provider, parent_state_root)? at line 119 is missing a trailing semicolon, while the equivalent call at line 84 has one. Both compile, but the inconsistency is worth fixing for readability.

4. Unused import: DatabaseStateRoot is imported at line 20 but not directly referenced in user code (it's a trait impl). Verify this is actually needed — if it's only providing the from_tx method on StateRoot, it should stay, but it may trigger an unused-import warning depending on compiler settings.

[claude]

Performance concern: StateRoot::from_tx(provider.tx_ref()).root() recomputes the state root by walking the trie cursors. This is significantly more expensive than the previous check which was a simple block hash lookup.

This function is called once per storage proof target (in a parallel iterator at line 84) plus once for account proofs (line 119). For blocks with many touched accounts, this could add substantial overhead.

Is this intentional? If the goal is just to verify the DB hasn't been updated underneath, could you read the root node directly from the account trie table instead of recomputing it?

[dvush]

Is this intentional?
Yes it is, is once per database transaction.