cmd/utils/app: import all RLP files in a single process

[yperbasis]

Problem

The nightly ethpandaops/hive-tests legacy (ethereum/consensus) suite fails 9 tests against erigon main — all variants of ForkStressTest (×5) and walletReorganizeOwners (×4). They are timeouts, not consensus failures: each ran ~180.5s (hive's --client.checktimelimit=180s) and was killed mid-import with no errors.

Root cause: the consensus simulator feeds one RLP block per file (/blocks/NNNN.rlp, no chain.rlp), and erigon's hive entrypoint invoked erigon import once per file. Each invocation is a full node startup (DB open, P2P networking, private gRPC RPC, MCP server) costing ~0.66s before a single block is read. Reorg-stress tests have hundreds of blocks (ForkStressTest = 303, walletReorganizeOwners = 259), so 303 × 0.66s ≈ 200s > 180s → the client never becomes RPC-ready in time. The other 32 607 tests have few blocks and pass.

Fix

importChain documented multi-file support (<filename> (<filename 2> ... <filename N>)) but only imported cliCtx.Args().First(). It now imports every file in one node process via importFiles, tolerating per-file failures when several files are given (fatal for a single file) — matching the documented contract and preserving the InvalidBlocks-test behavior the old per-process loop provided. The MCP server is also force-disabled for the one-shot import, alongside the existing NAT/downloader/external-consensus disables.

This is the binary half of the fix. The load-bearing companion batches the per-file loop in hive's clients/erigon/erigon.sh into one erigon import invocation (separate PR to ethereum/hive); both are needed to clear the suite.

Validation

Reconstructed the exact hive scenario from the local legacy-tests fixtures + client logs and ran the rebuilt binary:

Test	Blocks	Processes	Time	Resulting head
ForkStressTest_Frontier (12-fork reorg)	303	1	<1s	0xc14ab05e… ✅ = lastblockhash
walletReorganizeOwners_Homestead	259	1	<1s	0xda82854c… ✅ = lastblockhash
ForkStressTest + injected malformed files	—	1	—	bad files skipped (no panic), head still 0xc14ab05e… ✅

The in-process side-chain/reorg logic yields the correct canonical head ~200× faster. This mirrors the native testutil.BlockTest harness, which already loops InsertChain per block in a single process for these same tests.

• make lint → 0 issues
• make erigon integration build clean
• New unit test import_cmd_test.go (TDD, Red→Green): all-files-imported + single-file-fatal + multi-file-tolerant

🤖 Generated with Claude Code

[yperbasis]

Closing as a duplicate of #21513, which is the canonical PR for this fix (ethereum/hive#1519 references #21513 by number as its dependency). #21513 is functionally equivalent. The only delta here was also force-disabling the MCP server during one-shot import — happy to fold that into #21513 separately if wanted.