ops: record artifact hashes, RAM limit, and resume regression

[AuraIis]

Artifact identity completed

Records SHA-256 for:

• all three German training bins
• selected step_60000.pt, sidecar and run manifest
• related but explicitly non-selected best.pt and sidecar

The selected launch base remains step_60000.pt; best.pt is not treated as equivalent.

Dedup capacity decision

• measured 1,447.6 bytes/doc over a combined 1M-document sample
• 30,441,496 reference documents total
• projected LSH RSS 44.1 GB against 49 GB available
• exact-hash set adds an unmeasured ~1.5–3 GB
• combined index rejected
• required execution: sequential single-reference passes, expected 6–12 GB peak/pass

The larger source pools are recorded as conservative overlap references without claiming that each was the exact tokenization source.

Resume evidence correction

Replaces the earlier unreproducible “bit-identical” observation with the clean-checkout result:

• scheduler LR identical
• eval loss nearly identical
• training loss diverges immediately
• root cause: loader-owned language/shuffle RNG streams restart from seed

Launch remains blocked until #14 is merged and the clean-checkout resume smoke passes with the fix.