Rates Engine v0.5.0-rc.75

[v0.5.0-rc.75] — 2026-05-24

Changed

• verify-archive-tier-a.service switched to Type=notify +
  WatchdogSec=1h (#62). Replaces the Type=oneshot + fixed
  TimeoutStartSec pair. The binary signals READY=1 at start and
  WATCHDOG=1 every 30 s for the rest of its life; systemd SIGTERMs
  only on real silence (binary hung / crashed / dead-locked), not on
  a guessed wall-clock cap. The walk can take 25h+ on single-chunk
  serial without anyone re-tuning a TimeoutStartSec; as long as
  it's making progress, the watchdog stays satisfied. The previous
  17h → 36h TimeoutStartSec raise from earlier in this Unreleased
  block is superseded by this change. sd_notify is a no-op when
  $NOTIFY_SOCKET isn't set, so manual ratesengine-ops verify- archive invocations from a shell are unaffected. Adds
  github.com/coreos/go-systemd/v22 as a Go dependency.

Added

• verify-archive per-chunk resume across restarts (#62). The
  state file previously was written only on clean exit ("at the end
  on success" per the doc comment) — a SIGTERM during a bootstrap
  walk discarded everything verified so far, so
  -from-last-verified on the next fire restarted from genesis.
  Yesterday's #62 walk burned 17h05m of clean work (ledger 2 →
  40.5M, every line verified) before the timeout SIGTERMed it; with
  state writing only on completion, that progress couldn't be
  recovered.
  • The state file now carries a per-tier InProgress section with
    the original run plan (from / to / workers / chunks) plus a
    Done flag per chunk. As each chunk's walker completes
    successfully, the orchestrator marks that chunk Done and
    rewrites the state file atomically. SIGTERM at any point leaves
    a coherent record of which chunks finished.
  • On the next fire, verify-archive reads the prior InProgress.
    If the run plan still matches (same from / to / workers /
    chunk-count), only the chunks that aren't Done get re-issued —
    completed chunks are skipped. A plan mismatch (operator changed
    flags, tip moved) ignores the prior state cleanly and starts
    fresh, with a log line naming the difference.
  • Worst case: a SIGTERM mid-walk loses one chunk's in-flight work
    (~1/12 of the bootstrap, not the whole thing). Mid-chunk
    resumption would require persisting a per-chunk -resume-from- hash anchor — deferred until the 1/12-loss proves too costly.
  • The end-of-walk updateTierState unconditionally clears
    InProgress now, including on no-advance success (every chunk
    was already Done from the prior run). Legacy state files
    without the in_progress key parse cleanly as nil.