Stage A — Source readiness (done)

Stage B — Theory extraction (done)

Define 3 benchmark tasks spanning:
- frozen-kernel-favoring setting,
- feature-learning-critical setting,
- depth-sensitive residual setting.
Pre-register measurable hypotheses per task.
For each task, include at least one disconfirming criterion.

Implemented baseline metrics script: experiments/src/tp_bridge_metrics.py
Added synthetic trace fixture: experiments/src/tp_bridge_example_trace.json
Produced example output: experiments/results/tp_bridge/example_metrics.json
Emit run metadata:
- parametrization class,
- optimizer class,
- width/depth vectors,
- LR schedule and batch settings.
Emit dynamics observables from real training run scaffold (run_tp_bridge_toy_mlp_purepy.py):
- update norms per layer,
- representation drift,
- kernel/tangent drift,
- transfer regret.
Add spectral diagnostics from real training runs (currently optional; requires eigenspectrum source).

If NTK drift is high in claimed kernel regime, reject kernel-regime claim.
If µTransfer fails under verified µP scaling, flag mismatch between assumptions and implementation.
If depth transfer fails in depth-µP candidate setup, classify likely block-structure limitation.

Derive preliminary proxies for:
- generalized mass tensor behavior,
- coupling tensor behavior,
- representational cost vs predictive gain.
Report where results are:
- relabeling,
- structural consequence,
- genuinely novel.

This phase is done when:

all TP-series source artifacts are locally reproducible,
theorem-assumption-test mapping exists,
notation bridge and non-equivalence boundaries are explicit,
experiment checklist is concrete enough to implement without re-reading all papers.