add repro for cu issue (do not merge)

[kevinheavey]

@Aursen sharing progress so far

[procdump]

Two more eyes here on the issue @Aursen , @kevinheavey - anza-xyz/agave#10659. Do you plan to switch LiteSVM to BPFLoaderUpgradeable?

[procdump]

@kevinheavey , @Aursen - anza-xyz/agave#10659 (comment)

[Aursen]

@kevinheavey , @Aursen - anza-xyz/agave#10659 (comment)

Thank you for the help

[cds-amal]

Hey 👋 @kevinheavey & @Aursen ,

I was looking for a way to recover the feature set as of a given epoch (for replaying older transactions) and found this PR/Issue.

Have you considered the feature set a transaction runs under? Each activation can meaningfully change a transaction's CU, and mainnet_feature_set() means mainnet today, not mainnet at the snapshot's slot, so any capture that predates an activation replays under rules the chain wasn't using yet.

This snapshot appears to predate two of them:

• SIMD-0266 (Efficient Token program, replace_spl_token_with_p_token):
  the saved simulation logs show classic SPL token costs (TransferChecked at
  6238 CU), so leaving the swap active would replay Tokenkeg on a different
  implementation entirely.
• SIMD-0339 (Increase CPI Account Infos Limit,
  increase_cpi_account_info_limit): it changes CPI translation
  charging in solana-program-runtime/src/cpi.rs from actual-data-length to
  fixed-size accounting, which would surface as a small per-CPI difference
  in the caller's meter. Might be worth checking whether deactivating it
  closes part of the residual here.

The nice property is that the snapshot already witnesses its own epoch: the saved logs price CPIs the pre-0339 way, no activation-date lookup required. Which suggests the feature set belongs in the snapshot alongside the accounts and programs, recorded at capture time (or at least settable per replay), so a saved repro stays true indefinitely instead of drifting as mainnet activates features.

And if loader shape and feature epoch were both controlled, whatever remains would point at separate axes worth eliminating in turn: builtin instruction base costs, and bundled program builds (token-2022 here) differing from the bytes deployed at the captured slot.