fix: make checkpoint resume explicit and stop claiming optimizer state

[Ayyankhan101]

The last piece of the training path that was written, claimed, and never checked.

Resume was implicit, and could silently do nothing

train() loaded any checkpoint it found with no flag and no way to opt out, restoring the
step counter along with the weights. So re-running glm-train --steps 80 after an 80-step
run restored step = 80, fell straight through the loop, and printed what looked like a
successful run having trained nothing.

Resume is now opt-in via --resume:

$ glm-train --steps 40                     # second time, no --resume
Note: checkpoints exist in "glm_checkpoint" but --resume was not passed;
starting from step 0 and overwriting.
Starting training from step 0

$ glm-train --steps 40 --resume            # already at the limit
Resumed from step 40 (optimizer moments restart cold)
Error: nothing to do: resumed at step 40 but max_steps is 40.
Pass a larger --steps to continue training.        # exit 1

$ glm-train --steps 80 --resume            # continues
Resumed from step 40 (optimizer moments restart cold)
Starting training from step 40
Step 45: loss = 5.1103 (avg = 5.2861), lr = 7.04e-4

save_optimizer_state did not save optimizer state

It wrote {"step": N} — 16 bytes, no AdamW moments — while the README claimed "model +
optimizer state", and load_checkpoint read that file into an empty if let whose body
was a comment admitting nothing happened.

candle keeps first_moment / second_moment in a private VarAdamW with no accessor, so
honouring the claim means writing and maintaining our own optimizer. Not worth it for a CPU
playground with short runs, so the claim goes instead: the config field, the
optimizer_step_*.json file, the empty if let, and the README line are removed. What
resume does restore — weights, step counter, LR schedule position — is now documented,
along with the fact that Adam restarts cold and the loss briefly rises.

Also

save_checkpoint fell back to the literal name "unknown" when the parameter and name
lists disagreed in length. That collides in the safetensors map and silently drops every
parameter after the first. It now errors.

Verification

The check that matters is continuity across a resume. Last loss before: 5.0751 at step
40. First after: 5.1103 at step 45, with the learning rate picking up at the right point
on the cosine schedule. A cold start would have been back near 10.8, so the weights
genuinely restored.

New tests: a checkpoint round trip asserting every parameter matches after save and load,
and a check that the parameter and name lists agree in length and carry no duplicates —
the positional pairing is only safe while both hold.

104 tests pass
cargo clippy --all-targets --all-features -- -D warnings   exit 0
cargo fmt --all -- --check                                  clean

[git-mind-pr-guardian]

GitMind PR Review

{'pr_url': '#17', 'repo': 'Ayyankhan101/Transformer-In-Rust', 'pr_num': '17', 'title': 'fix: make checkpoint resume explicit and stop claiming optimizer state', 'author': 'Ayyankhan101', 'additions': 206, 'deletions': 87, 'changed_files': 9, 'review': '## Summary\nAutomatic review could not be generated — please review the diff manually.\n\n## Issues\nNo significant issues found.\n\n## Verdict\n💬 NEEDS DISCUSSION\nThe automated review failed; human review required.'}

---

Auto-generated by GitMind — AI-powered code analysis.