feat(cadecon): kernel-RMSE convergence metric + reworked asymptote dashboard

[daharoni]

Summary

Replaces CaDecon's iterative convergence stopping criterion with a peak-normalized RMSE between successive iterations' kernels, and reworks the Asymptote dashboard so the two signals we actually watch for convergence read as a matched pair.

Previously convergence was gated on the max relative change of the kernel's (t_peak, FWHM). That over-penalized t_peak jitter on the poorly-constrained rising edge — a few-ms wobble read as ~8% relative change and reset the patience streak, delaying convergence over a change that barely altered the kernel. A whole-waveform RMSE weights each parameter's change by how much it actually moves the kernel (that same wobble now reads ~1%).

Metric

• Computed on the analytic parametric kernel (peak = 1 by construction — no extra renormalization), not the un-normalized free-form h_free.
• Both kernels sampled on a shared native-fs grid over [0, 5·max(τ_decay)] (the max never truncates a genuine tail difference), floored to 24 samples so a low-fs / fast kernel doesn't collapse to a handful of points.
• convergenceTol default 0.02 → 0.005 (now a fraction of peak, ≈ a 2% τ change). Kernel RMSE remains the sole gate; the final kernel is still the median shape of the trailing window.
• Activity-stability metric now uses a symmetric max(‖s_new‖, ‖s_old‖) denominator so a collapsing cell can't blow up the ratio.

Dashboard

Asymptote tab reworked into two rows:

• Top (log-y, decay-toward-0): Kernel change (RMSE) with a reactive tolerance band + patience highlight; Activity change with a marker showing both the gate iteration and the +1 iteration where activity actually settles (structural lag — spikes are inferred from the prior kernel).
• Bottom (linear, rise-to-plateau): kernel-fit R², reconstruction PVE.

Code quality

• Convergence metric lives in @calab/compute (kernel-math.ts), sharing a sampleBiexp core with computeKernel, with direct unit tests.
• Chart markers share a drawVerticalMarker primitive; the two trend panels share a TrendChart shell.
• Docs, README, and tutorial steps updated to describe the RMSE criterion.

Testing

• Typecheck passes across all apps (cadecon, catune, carank, admin) and packages.
• Lint + Prettier clean.
• Tests pass: compute 81 (incl. new kernelShapeRmse cases), cadecon 33, catune 26.
• Manually exercised in the running app; convergence and activity now flatten within one iteration of each other with the gate/settles markers straddling.

🤖 Generated with Claude Code