generate_bouquet leaks coil regularization + drift bounds onto the TokaMaker object (needs teardown)

[d-burg]

generate_bouquet installs a strong coil soft-regularization (weight=1e4 swap) and _coil_drift_bounds during draw generation. These live on the TokaMaker object — not the equilibrium — so mygs.replace_eq() does not undo them and they outlive the call.

Measured consequence (2026-08-04, fix/rng-seeding-and-anchor-ip test development): a second in-process run on the same solver lands coil currents ~1.4e-3 apart and l_i ~0.6% apart versus a fresh process. The branch's reproducibility tests work around this with subprocess isolation.

Proposed fix: a teardown (context manager or explicit restore) that snapshots the coil reg/bounds state on entry and restores it on exit of generate_bouquet, so sequential in-process runs (run_slices, notebooks, future solve_point scans) are state-clean. A regression test = two sequential in-process σ=0 runs must be bit-identical.

[d-burg]

Location + evidence pointers (2026-08-04 session): the strong soft-reg (weight=1e4) swap and _coil_drift_bounds stash happen in the per-draw path of bouquet/TokaMaker_interface.py (the l_i-loop/homotopy machinery around the state-anchor; grep _coil_drift_bounds). They live on the TokaMaker object, so mygs.replace_eq() does not undo them — that asymmetry is the leak. Measured: second in-process run lands coil currents ~1.4e-3 apart, l_i ~0.6%. Related but distinct: repeated anchor+SWB WITHOUT an equilibrium restore drifts ~1% in l_i (state evolution, not a leak — it initially masqueraded as SWB non-determinism; see the retraction in issue #16's proposal doc). Affected use patterns: run_slices, sequential notebook runs, and the proposed solve_point() beta scans (#16) — all sequential in-process solver reuse. Regression test spec: two sequential in-process sigma=0 runs must be bit-identical; current tests in tests/test_seeded_reproducibility.py dodge this via subprocess isolation and can be tightened once fixed.