Show a big roll's tail faster as its dice are read

[drehtuer]

Owner feedback: 100d4 takes a long time to settle. Of the options offered (faster tail, fewer cocked d4s, chase the rare give-up), the owner chose A, the faster tail.

Measured on the Pixel 10a (tools/harness.sh -n 200 -c 100 -s d4, main at #372):

• median settle 1.29 s, p99 3.09 s of simulated time;
• 1 roll in 200 gave up;
• 3.07 % of dice re-thrown.

At the watched pace of 0.4, that is 3.2 s and 7.7 s to watch. Nearly all of the tail is a few dice rocking among ninety still ones.

Change (decision 88): for a roll of 10 dice or more, RollPace.paceFor climbs from the watched 0.4 to real speed as the share of unread dice falls:

unread share	pace
≥ 25 %	0.4 (watched)
25 % → 10 %	straight line
≤ 10 %	1.0 (real time)

• The climb is gradual, so it doesn't read as a dropped frame.
• It never passes real speed, so the last dice still move like dice.
• It only decides when steps are taken: faces, fairness and goldens are unchanged.
• Rolls under 10 dice keep 0.4 to the end, because there the last die tumbling is the roll.
• TrayLoop passes unsettled.size and the dice count to secondsFor.

Docs:

• docs/physics-and-rendering.md, "The simulation clock": replaces the old "one constant, no easing" paragraph.
• Decision 88, which refines 65.
• TODO 5.3: the 100d4 item carries today's phone numbers and names cocked d4s as the next step.
• TODO 5.6: a phone check for the faster tail.

Tests:

• RollPaceTest: small rolls unaffected, the quarter and tenth bounds, a monotonic line within [0.4, 1.0], a hand or a long roll still wins.
• TrayLoopTest: a 100-die roll with 5 still moving is advanced at real time through the real frame path.
• :simulation:api and :render:filament unit tests, ktlint, detekt and the docs checks pass.

Judge on the phone: does the end of 100d4 and 40d6 speed up smoothly and feel short enough now, while 1d20 and 5d6 still feel unhurried?

Based on main (#372).

🤖 Generated with Claude Code

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
93.8% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud