Reject stale completed one-head nanoGPT runs

[charlesmartin14]

Root cause

The one-head nanoGPT runner treated the existence of run_complete.json as sufficient proof that a result could be reused. That shortcut ran before the current prepared-data identity and protocol fingerprint were calculated. run_is_complete() likewise checked only completed: true.

A stale directory from an older protocol, different prepared corpus, changed optimizer profile, changed model/training configuration, or missing terminal artifact could therefore be silently accepted.

Changes

• add a strict completed-run validator for the one-head nanoGPT suite
• load and verify prepared-data metadata before deciding whether a completed run is reusable
• calculate the current optimizer profile, training horizon, and protocol fingerprint before the completion shortcut
• validate completion and manifest fingerprint, optimizer, seed, and total-step identity
• validate final and validation-selected test-result steps
• validate metrics and epoch-metrics coverage, monitoring-only test policy, spectral coverage, six-matrix inventory, and the exact epoch-checkpoint paths recorded by the run
• validate latest, final, and validation-best checkpoint fingerprints, optimizer/seed identity, steps, and best-validation state before training skips a completed run
• make run_is_complete() use artifact validation instead of trusting a Boolean marker
• add regression tests for valid reuse, protocol drift, missing artifacts, and stale checkpoint fingerprints

Behavior

A compatible and complete result is reused without initializing the training device or model. A stale, incomplete, corrupt, or ambiguous result fails explicitly and instructs the caller to use a new result directory or rerun with explicit overwrite. Existing artifacts are never silently replaced.

Validation

• git diff --check
• python -m compileall -q baseline/nanogpt_one_head/src baseline/nanogpt_one_head/tests
• pytest -q baseline/nanogpt_one_head/tests — 17 passed
• pip check — no broken requirements

No optimizer mathematics, hyperparameters, model architecture, notebook source, committed output notebook, or experiment result was changed. This PR is independent of PR #40.