Restoration without git history

The hardest restore is not a clean clone. It is a machine image or archive where the code still exists, some receipts still exist, but the Git history that used to explain the tree is gone.

In that situation the only honest workflow is reconstructive: choose the most plausible upstream base, replay the narrow local patch surface, and then prove the rebuilt tree still behaves like the lane you meant to recover. That is why this short note sits next to Restoring a Megatron training tree without git history, Checkpoint format and resume, and How we keep a patch lane.

For first touch:

• The upstream base is the closest verifiable upstream snapshot that explains the surviving tree with the smallest local delta.
• The patch surface is the bounded set of local seams still required on top of that base.
• The canary is the smallest launch or resume proof that shows the rebuilt tree still behaves like the intended lane.

The quickest public-safe starter kit is:

• NAM56R Megatron plan sample
• NAM56R runtime patch surface sample
• NAM56R launch contract sample

Together they separate base plan, local seam inventory, and proof launch.

What the workflow preserves

The goal is not to reconstruct every historical commit. The goal is to recover a trainingQuick term guideTrainingA grounded walkthrough of how the project approaches small-language-model training: explicit stack specs, memory-first patches, hybrid blocks, and…GroundingSLM training in MegaCpp: what the stack optimizes for and what stays explicit Training speed anatomy on H200 tree whose runtimeQuick term guideRuntime boundariesWhy MegaCpp pays first-compile and recompile costs in exchange for steady-state throughput, and the operational rules that keep torch.compile,…GroundingThe Compile-Time Tax We Accept for Runtime Speed Regional compile without losing the plot is explainable again.

That means preserving three things:

• a plausible upstream base
• a narrow replayable patch layer
• a verification path that proves the rebuilt tree still behaves correctly

That target is narrower than full archaeology and more useful for recovery. It matches the evidence-first posture in Checkpoint format and resume: restore the executable contract first, then widen the tree only if the verified lane still needs more context.

The public-safe sequence

1. Reconstruct the likely upstream-shaped execution plan.
2. Mark the local seams that still sit above that base.
3. Rebuild the launch contract that made the original lane comparable.
4. Run the smallest canary that still exercises the restored path.

The checked-in examples above are small on purpose because they keep each part of that sequence inspectable instead of hiding it inside one opaque launcher or archive.

The practical way to keep that scope honest is to inventory three separate contracts instead of diffing everything at once. The plan receipt says what upstream-shaped model you think you have, the runtimeQuick term guideRuntime boundariesWhy MegaCpp pays first-compile and recompile costs in exchange for steady-state throughput, and the operational rules that keep torch.compile,…GroundingThe Compile-Time Tax We Accept for Runtime Speed Regional compile without losing the plot patch-surface receipt lists only the live seams that still sit above that base, and the launch contract records the fixed policy that has to surround the generated args. When those three surfaces agree, you have a restoration story. When they do not, you usually have an archive that still runs but no longer explains itself.