[huangminghuang]

Summary

• Report a missing state database version header after interrupted initialization with a one-time --delete-all-blocks recovery instruction.
• Report an empty active blocks.log with its path and the same recovery instruction, while retaining archived-log priority and rejection of malformed logs.

Why

WIRE-420: a failed first launch can leave a headerless state database and a zero-byte block log. The previous errors suggested an incompatible snapshot or an unreadable file without explaining the recovery. Writing the database header during index registration would also let a partially loaded snapshot appear initialized on a later launch. Keep header creation at genesis or snapshot startup and fail with actionable diagnostics when initialization stopped earlier.

--delete-all-blocks clears the state database and block log while leaving finalizers/ intact.

Integration with current master

• Merged master at 6ded364fde2a86ddd2cb3d2d3000dd68e8e0e2a5 in merge commit 505743961f94b51252e8d3d50feb08e970bee05b.
• Resolved the overlap in restart_chain_tests.cpp by retaining both WIRE-420 diagnostic cases and the WIRE-389 transaction-dedup persistence cases. The PR's net diff remains the four intended controller, block-log, and test files.

Validation

• On the merged head, rebuilt unit_test with Clang 18 Release and ran the restart_chain_tests and block_log_extraction_tests CTest suites under --sys-vm-jit: both passed.
• Before the base merge, focused restart_chain_tests and block_log_extraction_tests under --sys-vm-jit passed (12 cases, 59 assertions), including both exact recovery messages. Focused chain_plugin_tests passed (5 cases, 123 assertions).

[heifner]

I'd rather not make interrupted initialization restartable. A first launch that fails before chain startup leaves nothing worth keeping: an empty state database and a 0-byte blocks.log. Rerunning once with --delete-all-blocks is the recovery. Creating the header early also adds a new state: a snapshot load that aborts partway now leaves a header at revision 0, and a later genesis start initializes over the partial rows instead of refusing.

What needs fixing is the error messages. After a first launch fails in a later plugin's init:

• The relaunch fails with "state database version pre-dates versioning, please restore from a compatible snapshot or replay!", which points the wrong way.
• Removing only state/ (which the --snapshot error tells you to do) then fails on the empty blocks.log: "cfile: …/blocks.log unable to read 4 bytes; only read 0, eof: true".

Could you replace the header and block-log changes with two clear errors?

1. Missing header (validate_db_version): "State database has no version header: the launch that created it did not complete startup. Rerun once with --delete-all-blocks."
2. Empty active blocks.log (extract_block_log_preamble): a block_log_exception with "<path> is empty: the launch that created it did not complete startup. Rerun once with --delete-all-blocks."

--delete-all-blocks keeps finalizers/, and the --snapshot-endpoint error already gives the same advice. A test for each message is enough.