b70-optimization-lab

Laguna USB-backup-only NVMe migration

The active Laguna S 2.1 target and DFlash draft are now on the internal NVMe:

/mnt/fast-ai/llm-models/laguna-s-2.1/int4
/mnt/fast-ai/llm-models/laguna-s-2.1/dflash-int4

The external Corsair ntfs3 copy remains intact, but it is backup-only. Live model reads, compiler/model caches, temporary files, logs, run roots, and recovery evidence must not use /media/steve/CorsairExternal. Frozen historical notes continue to name their original evidence paths and must not be rewritten as if those runs occurred on NVMe.

Copy verification

The copy used an isolated .incoming directory. It was not promoted to the final model path until independent relative-path SHA-256 manifests from the USB source and NVMe destination compared byte-for-byte equal.

Item Files Logical bytes
INT4 target 100 72,021,173,058
INT4 DFlash draft 18 2,229,975,225
Total 118 74,251,148,283

Both retained manifests have SHA-256:

45aa105ef4eceaf05cad33012e0752369f77cbbd76f2213ccfe0ce130fa6c0ac

They are stored at:

/mnt/fast-ai/llm-models/laguna-s-2.1/.verification/source-files.sha256
/mnt/fast-ai/llm-models/laguna-s-2.1/.verification/nvme-files.sha256

The teacher, four-card formal W1 results, aggregate formal/counter summaries, real M8 timing fixture, and peer binary were separately mirrored under:

/mnt/fast-ai/llm-optimization-artifacts/laguna-s-2.1

All nine local hashes match the already pinned source hashes.

Space reclamation

The NVMe initially had only 7.7 GiB available. Exactly six Hugging Face cache roots were removed, reclaiming 132,578,820,096 allocated bytes (123.47 GiB). They belonged only to the closed Qwen 3.6 27B lane or the blocked native-XPU Qwen 3.6 35B FP8 diagnostic. Their repository identities and revisions remain recorded in Git and in the structured migration packet, so they can be downloaded again.

No DeepSeek, Gemma, MiniMax, source, benchmark-result, or unique experiment evidence directory was removed. After the verified Laguna copy, ext4 had 66,551,144,448 bytes available.

Benchmark disposition

This was storage-only work. No XPU command, model service, model generation, candidate, or benchmark ran. The current boot remains rejected because the earlier direct ntfs3 evidence write produced kernel taint 640. The local runner/gate migration must be committed, then the host must reboot before any GPU work. A distinct no-generation recovery gate must pass before recovery A1, which remains the first subsequent model generation.

Structured packet: data/laguna-s-2.1-nvme-model-migration-20260723.json.