The remote tracker increments imm_words/imm_kanji.frequency live, so a
snapshot holding a stale ACTIVE session (skipped by the merge) has that
session's partial occurrences baked into frequency. A word new to the
local DB adopted that full total, then received the same occurrences
again when the session finalized and merged on a later sync.
Subtract active-session occurrence counts when adopting a new row; they
are re-added exactly once when the session completes. Existing rows were
already safe (isNew guard in addWordOccurrences/addKanjiOccurrences).
Verified red/green: the new two-sync regression test fails without the
subtraction (frequency 9 instead of 5).
The engine now executes only inside the app (libsql): subminer sync
rebuilds the equivalent --sync-cli argv and spawns the discovered app
binary with the terminal attached (stdin for ssh prompts, raw NDJSON
stdout, exit-code passthrough). The bun:sqlite driver binding and the
launcher-side engine shims are gone; flow tests moved to
src/core/services/stats-sync/sync-flow.test.ts, ssh tests to src, and
the merge suite now runs through the libsql driver the app ships.
sudacode
