Feature-flag / capability registry with scoped sources (directory, harness, etc.)

[MasonRhodesDev]

We keep accumulating places where UI/behavior should be conditional on what the thing is: harnesses (#29 — what does this agent support?), directories (does this project have git?), and in the future per-MCP, per-user, per-build-mode, etc.

Build a single feature-flag/capability registry with scoped sources so every call site asks the same question:

features.is_enabled("pull_requests", &scope)

…where scope is composed of contributors (directory detectors, harness capability bits, user prefs, experimental toggles). Contributors resolve in a defined order; the resolved value is cached and invalidated when its sources change (e.g. .git appears or disappears). Flags carry metadata (label, description, default, requires) so a debug/developer panel can list what's active and why.

Payoff:

• Unify agent harnesses behind a common capability interface #29 (harness trait) gets a real home for its capability bits.
• Directory-capability work (filed separately) slots straight in.
• Future dynamic UI (per-harness cheatsheets, per-project toolbars) stays declarative.
• Experiments / feature previews can be toggled without touching call sites.

Out of scope here: the exact API shape — design TBD.