feat: add mobility context for instrument applicability

[murdercode]

Problem

Many Apple Health-derived metrics are produced by step and gait detection. When that detection cannot run, the host does not receive a low value — it receives a value that does not describe the person. Wearables have been reported as recording zero steps for someone walking with a walker. The calculator had no way to tell that apart from genuine inactivity, so a wheelchair user was scored on steps they cannot take.

What this adds

FitnessAgeMobilityContext on FitnessAgeProfile, declaring which movement instruments can be observed:

Context	Covers	Removes
ambulatory (default)	unaided walking, cane	nothing
assistedAmbulation	walker, crutches	step-derived and gait-derived instruments
nonAmbulatory	wheeled mobility	the above, plus stand_hours

Inapplicable IDs are unioned with disabledMetricIds before scoring, so removal follows the already-documented path: the field is cleared, the domain divides by the observed local weights that remain, and data quality averages over observed components only.

This is a measurement-applicability input, not a diagnosis or clinical classification. No threshold, curve, local weight, domain weight, or confidence rule changed.

Where the boundaries come from

Boundaries follow measurement literature rather than a diagnosis list:

• Step counting is described as applicable only to ambulatory populations, degrading further under irregular gait — Suzuki et al., JMIR Formative Research 2025.
• Weight-bearing aids break step detection rather than adding noise: poor validity with gait aids, near-zero counts during walker use — Kooner et al., J Orthop Surg Res 2024. Wrist error with a walker 31.2% — Jaworski et al., IJERPH 2025.
• Cane use sits on the other side: ~2% wrist error, which is noise, not inapplicability — same study. Cane users stay ambulatory.
• Activity targets are not context-specific; WHO 2020 added recommendations for people living with disability, so activeEnergy and exerciseTime stay the applicable instruments instead of being replaced — Bull et al., BJSM 2020.

Every anchor is in the methodology source table.

Deliberately out of scope

Conditions where the instrument works but the expected value is debated (fatigue-related, mental-health-related). Step count is measurable there; changing its target without an evidence anchor would be a scoring change presented as an applicability change. That needs an Algorithm RFC.

Compatibility

No result drift. mobilityContext defaults to ambulatory, which removes nothing; encoded profiles omitting the field decode to ambulatory. Golden fixtures unchanged.

Verification

41 tests pass, 9 new. Two carry the weight:

• golden parity intact — no drift for existing integrations
• inapplicableInstrumentsAreRemovedRatherThanScoredAsZero — a context-driven removal produces byte-identical results to an input that never supplied those metrics. That is the invariant that makes this correct rather than charitable.