ops: record measured server, tokenizer, data and FineWeb2 probe state

[AuraIis]

Measured machine-lane values

• Python 3.11.12, torch 2.7.0+cu128, Triton 3.3.0, CUDA build 12.8
• NVIDIA RTX PRO 5000 Blackwell, sm_120, 48 GB
• required compatibility override TRITON_OVERRIDE_ARCH=sm89
• exact tokenizer path, size, 200k vocabulary and full SHA-256
• exact training-bin paths, byte sizes and token counts
• exact 7,461,038,089-token distinct German inventory
• source paths/sizes and the completed FineWeb2-DE source SHA-256
• raw FineWeb2-HQ probe report: 45.11% dropped against german_fresh on shard 1/570

Evidence boundaries

The probe is recorded as one shard against one reference using the legacy report. Its single-reference gross projection is ~15.64B retained tokens; this is not presented as the all-reference full-pull result.

The full pull is approved after disk preflight, but promotion remains blocked until the main-branch dedup script produces a complete all-reference report.

Remaining launch blockers

• exact launch commit from the clean server checkout
• basis-checkpoint file manifest hash
• hashes of the actual tokenized training bins
• completed four-reference dedup manifest
• full FineWeb2-HQ all-reference dedup report

Tests lock the delivered measurements and prevent the probe from being treated as full evidence.