The in-process registry can be snapshotted to disk. Persistence is opt-in via SPIKEFORGE_METRICS_PERSIST; with it unset, flush() is a no-op and behaviour is unchanged. Snapshots (MetricSnapshot) are written under METRICS_DIR (SPIKEFORGE_METRICS_DIR, default <DATA_DIR>/metrics). system_stats gains additive metrics_persisted and metrics_last_flush keys; existing keys are untouched.

TrainConfig.tracking selects tensorboard (extra tracking) or wandb (extra tracking-wandb), default null. The local manifest is always written first, so a tracker outage never loses a run; an absent backend becomes a recorded reason in the manifest's tracking block (sinks.describe()). Probes are isolated in tracking/sink_probe.py.

tracking.determinism.enable_deterministic() seeds Python/NumPy/torch and sets the deterministic-algorithm flags, returning a report of what it could and could not enforce; bit_exactness_check() reruns a fixture and reports exactness. TrainConfig.deterministic (default off) opts a run in, and the manifest gains an additive determinism block. This narrows the bit-exactness gap; it does not claim universal bit-exactness.

mkdocs.yml and scripts/build_docs.sh render plans/ and this README into a browsable Material site:

scripts/build_docs.sh          # build into build/docs
scripts/build_docs.sh --check  # fail on broken documentation links

The markdown remains authoritative; the site is a rendering of it. The docs extra provides MkDocs Material.