[xtofalex]

Why

A review snapshot was named by the sha256 of its naja-if dump. The dump records the producer's naja version, so the same netlist got a different id with each naja build, and threads written on one machine could show as other_snapshot on another.

What

• Snapshot id = NLUniverse.getContentHash() (naja #514, najaeda >= 0.7.35). The hash is the same after a naja-if reload, so a reopened snapshot resolves its threads exact. A fresh load of the HDL/Liberty sources gets the same id only when naja assigns the same ids (file order matters). That is accepted: reopen the snapshot instead.
• Dumped on first thread only, straight into the review directory: written to a staging directory and then renamed, so the first writer of an id wins.
• One design DB per review: opening a thread is refused while the universe holds several user DBs. The snapshot dumps only the top DB, so it would not reload under its own id.
• load_review_snapshot() now returns the loaded top design (it returned None), and it warns when the reloaded netlist doesn't hash to its folder name.
• Tests:
  • A stored snapshot reopened in a fresh process hashes to its name, and its thread resolves exact.
  • Opening a thread is refused when the universe holds several design DBs.
  • The anonymous-instances fixture now destroys the DB it creates.
• najaeda floor 0.7.35: protocol.py, pyproject.toml, the CI matrix, and the thirdparty/naja submodule (v0.7.35).
• Docs: docs/review_format.md (Snapshots section) and CLAUDE.md.

Tested

• pytest python/tests: 181 passed on najaeda 0.7.35.
• Native macOS build against naja v0.7.35, ctest: 244/244 passed.
• Linux native CI has not been run against the new submodule yet.

Known limits

• A naja-if reload still loses Liberty sequential modeling, so flops in a reopened snapshot draw as boxes. This is a naja issue.
• No test yet checks that the C++ and Python hashes match. It comes with native review support.