Chore/bench weights spec 17

[nol4lej]

Summary

Regenerates every versioned weight on the reference hardware (Hetzner CPX62, ubuntu-32gb-fsn1-1, 50/20) after spec 17, and fixes the benchmark runner. Before this fix, the runner could not complete on the CPX62.

Runner (scripts/benchmarks/run_benchmarks.sh)

• --no-median-slopes. In the 2606 CLI, the median-slopes analysis grows with repeat on any benchmark that has a component. It is only printed: weights come from min-squares, the default of --output-analysis. Even so, it pushed pallet_relayer and pallet_ismp_messaging past 31 GB, and they were OOM-killed on the CPX62. sha3fips was OOM-killed even on a 64 GB host.

  Peak memory with the flag, measured locally at 50/20:

  Benchmark	Peak RSS	Time
  relayer	0.43 GB	98 s
  ismp	2.1 GB	287 s
  sha3fips	2.3 GB	191 s
  set_allowed_selectors (5×20)	0.40 GB, was 3.8 GB	13 s, was 131 s

  Ruled out first: PoV mode, --no-storage-info, --no-verify, --disable-proof-recording, and T::X::get() ranges.

• Runtime-side weights are rewritten to the upstream trait. Applies to every file under template/runtime/src/weights/, today ismp_grandpa. The script drops the local trait WeightInfo and the impl for (), and points the impl at <pallet>::weights::WeightInfo.

  • Before, this was a manual step. Forgetting it broke the runtime build, and every later run of the script with it, since the script rebuilds first.
  • The run fails if the rewrite does not apply.

• Proof-size guard. After regenerating pallet_ismp_messaging, the script runs declared_proof_sizes_stay_within_a_sane_ceiling. It fails the run when the CLI writes a corrupted proof_size (polkadot-sdk#13066).

• Comments and the exit-137 message are updated: they still said "~15 GB" and "no flag makes that fit".

Weights (CPX62, 2026-10-06)

• Regenerated: shielded-pool, zk-verifier, relayer, validator-set, ismp-messaging, ismp_grandpa.

• relayer and ismp had not been regenerated since August and September, before the SDK 2606 migration.

• shield / shield_batch:

  Extrinsic	Weight	PoV
  shield	822 M ps	4707
  shield_batch, per entry	785 M ps	2475

  This is the pallet's own cost, mostly the Poseidon Merkle insert. Groth16 verification is not included: it is added separately through verification_weight (~11.7 G ps).

• Hand-corrected dispatch_get proof_size. The CLI wrote 8433417681689641984 against 283 measured bytes. The base is now 3748, the value --steps 3 and --steps 5 produce. Proof size is measured storage and does not depend on the host, and 3748 also equals dispatch_post's base on this run, since both walk RequestCommitments. The per-key slope (k × 1874) and every execution time are the CPX62's. The "Hand-corrected value" header is restored for dispatch_get.

Test plan

• Full run on the CPX62: [3/3] All 12 benchmark(s) completed successfully.
• cargo check -p orbinum-runtime
• cargo test: ismp-messaging 86, shielded-pool 390, zk-verifier 128, relayer 106, validator-set 36.
• No other proof_size of 8 digits or more in the 6 regenerated files.
• Runner locally at 50/20: relayer, the ismp group and sha3fips complete. The rewritten ismp_grandpa.rs compiles, and it is byte-identical to the hand-fixed one on a simulated generator output.
• Proof-size guard: passes on the committed weights, fails on an injected 8433417681689641984, and is skipped for other pallets.