Named visual-noise setting for the opt-in clone (ADR-0024 accepted)

[Telschow]

Stacked on #82 (it contains the ADR-0024 draft and the wider measurement); the diff against main shows both until #82 merges.

What

• ADR-0024 decision: options 1 and 2. The clone stays opt-in, `vision_enabled` stays False, and four times the measured visual noise is the supported assumption for a feature-tracking two-view front end.
• A named setting for it: `VISION_NOISE_PRESETS["measured-front-end"]` (4.0) in `navkit.estimators.eskf`; `estimator.vision_noise_preset` in a scenario file (only with `vision_model: clone`, not together with `vision_sigma_scale`); `RunOptions.vision_noise_preset` for recorded data (not together with `vision_noise_scale`).
• A test ties the constant to the four-times run of the wider measurement. No shipped case or default uses the preset.
• MODEL.md, the real-front-end page, status, B1 in CONSTRAINTS, ROADMAP and CHANGELOG updated. Also corrects a stale line in MODEL.md that said no benchmark case sets the scale keys.

Checks

Full suite (1316 passed), lint, types, links, strict docs build, SRS trace and metrics pass locally. Coverage 94.31%.

🤖 Generated with Claude Code