test(ai): add AI sanitizer and fallback invariants

[JSONbored]

Background

If optional AI is introduced, public/private safety cannot depend on model behavior. Sanitizer and fallback invariants need strong tests.

Goal

Centralize tests for optional AI output safety.

Current Behavior

Public-safety tests exist for deterministic outputs, but AI-specific fallback behavior needs explicit coverage before enabling AI summaries.

Desired Behavior

AI output is always filtered, unsafe output is blocked, and deterministic fallback works across public/private summary modes.

Implementation Requirements

• Add fixtures for unsafe AI output.
• Test public and private summary modes separately.
• Test quota/unavailable fallback.
• Ensure private facts are not rendered publicly.
• Keep deterministic signals untouched by AI failures.

Public/Private Output Boundaries

Public GitHub output must never include wallets, hotkeys, payout/reward estimates, raw trust scores, public score estimates, private reviewability, private scoreability context, or farming language. Private API/MCP/control-panel surfaces may show authenticated scoreability, blockers, projections, and evidence, but must not claim guaranteed payout outcomes.

Acceptance Criteria

• Unsafe generated output is blocked.
• Private facts never appear in public output.
• Quota/unavailable model falls back cleanly.
• Tests cover public and private summary modes.

Testing Requirements

• npm run test:ci must pass.
• Global coverage must remain at or above 97% for lines, statements, functions, and branches.
• Aim for 98%+ branch coverage locally.
• Add tests for every new branch, fallback path, sanitizer rule, and regression.
• Add invariant/property-style tests when behavior depends on sorting, gating, scoring, source-upload safety, public/private boundaries, or telemetry privacy.
• MCP/local tooling must prove source contents are not uploaded when local metadata is involved.
• Public GitHub output must be tested against forbidden language: wallet, hotkey, raw trust score, payout, reward estimate, farming, private reviewability, and public score estimate.