Add resumable Jacobian GPU workflow

[arthurpanhku]

What changed

• align the paper estimator with Anthropic's replicated-prompt, ordinary-VJP execution path
• add atomic checkpoint/resume with model, layer, parameter, and ordered-corpus fingerprint validation
• add a cloud-GPU fitter with preflight checks, WikiText streaming, smoke corpus, timing, peak-memory, and provenance reports
• allow the HF sidecar to load a fitted Jacobian while validating its companion metadata and tensor SHA-1
• add upstream reproduction, Colab/community capture, trace-gallery validation, contribution guidance, and CI
• update GPU runbooks, platform plan, README, and frontend dependency lockfile

Why

The earlier batched estimator used is_grads_batched=True, which is mathematically equivalent on the tiny validation model but does not match the upstream runtime path and can depend on vmap-compatible GPU backward kernels. Production fitting also had no safe recovery mechanism, making a long 7B run vulnerable to interruption and stale-checkpoint mixing.

Impact

J7Scope now has a reproducible path from mathematical validation through cloud-GPU benchmark and resumable fitting to sidecar loading and trace capture. The bundled corpus remains explicitly smoke-only; no real-GPU research result is claimed in this PR.

Validation

• python -m pytest -q — 24 passed
• python experiments/validate_jacobian.py --n-probes 1024 — exact/full-graph and replicated-batch/scalar VJP checks at zero relative Frobenius error
• python experiments/validate_trace_gallery.py results/traces — passed
• npm audit --audit-level=high — 0 vulnerabilities
• npm run build — passed (existing large-chunk warning only)
• git diff --check — passed

[socket-security]

Review the following changes in direct dependencies. Learn more about Socket for GitHub.

Diff	Package	Supply Chain Security	Vulnerability	Quality	Maintenance	License
	pypi/​datasets@​5.0.0

View full report