feat(sdk): conditioning / excitation / parallax monitors

[Ravenwater]

Report cond(H)/σ_min from solve_alignment (it silently adds a 1e-9 ridge and only fails on outright Cholesky breakdown — an under-excited solve yields a plausible-but-wrong gravity with no warning); add an excitation gate beside the gravity_tol bands in imu_init; surface KLT's already-computed min-eigenvalue + final photometric residual (discarded in klt.hpp) so the front end can weight/reject low-confidence tracks. Acceptance: a low-parallax/low-excitation window is flagged by a conditioning/excitation number, not by downstream divergence. Part of #261. See docs/assessments/vio-diagnostic-methodology.md §C.4.

[Ravenwater]

Triage (2026-06-14) — re-scope: partial coverage exists

Already present: the excitation/observability is characterized offline (eval/initialization_probe.hpp dynamic-excitation sweep), and S5 has a min-parallax gate (CameraUpdaterOptions::min_parallax_deg). Remaining delta — the runtime monitors this issue is actually about:

• report cond(H) / σ_min from solve_alignment (it silently adds a 1e-9 ridge and only fails on outright Cholesky breakdown — an under-excited solve yields a plausible-but-wrong gravity with no warning);
• an excitation gate beside the gravity_tol bands in imu_init;
• surface KLT's already-computed min-eigenvalue + final photometric residual (discarded in klt.hpp) so the front end can weight/reject low-confidence tracks.

Priority: P2.