[Feature]: [P3] session_fps samples idle (speed 0) and modal-frozen frames, diluting the fps signal on big towers

Deferred from the S5 gds-code-review (session_fps, #538). Edge Case Hunter, minor.

GameplaySession.noteFrame (src/analytics.ts) samples every foreground rendered frame, including frames where the sim is not advancing: game speed 0 (paused) and while a blocking modal freezes time. runFrame calls noteFrame as its first line, before the modal-freeze early return, and speed 0 still fires rAF -> tick -> render.

Why it matters: at speed 0 (or behind a modal) the per-frame CPU sim cost is skipped, so on a large tower, where the sim tick is the dominant per-frame cost, those idle frames run at a higher fps than active play. A player who parks at speed 0 or sits on a modal banks easy high-fps samples that pull p50/low toward the vsync ceiling and dilute the busy-simulation frames the metric exists to characterize. Big-tower sessions are exactly the Pixel 8a freeze scenario #538 targets, so the dilution is worst where the signal matters most.

Not fixed in the S5 PR because the clean fix couples the sampler to frame-loop state (game speed and modal-open), and gating sampling on "actively simulating" needs anchor-reset handling on every active<->idle transition (the wall-clock sampler must not charge the idle gap as one slow frame when active play resumes). That is more machinery than a minor bias warrants inline.

Options when picked up:

• Gate noteFrame on SPEEDS[app.speed] > 0 (and optionally modal-open), resetting the frame anchor whenever sampling pauses, so the reservoir reflects active render load only.
• Or keep sampling all foreground frames and document the metric as "all rendered frames, idle included" so the dilution is a known, accepted property.

Context: the S5 PR already fixed the larger correctness bug in the same area (the engine's 200ms frame-delta spike-clamp was inverting the worst-frame low signal); this deferral is the remaining measurement-validity refinement.

Start in src/analytics.ts with GameplaySession.noteFrame and runFrame, then trace app.speed, SPEEDS, modal state, and the frame-anchor handling. Decide whether idle and modal-frozen frames should be excluded or documented as included; done means the chosen behavior preserves valid active-play frame timing without charging idle gaps after sampling resumes.