test(clock): time-budget facts run on a fake clock (STORY-442, closes #723)

[genwave-radio]

Closes #723. PR-3 of the gh-#777 epic (STORY-442).

⏱️ What

• Microsoft's FakeTimeProvider replaces the two hand-rolled fakes in Orchestration.Tests and Tts.Tests (package added with dotnet add package; 34 consumers; Story241 sets its zone with SetLocalTimeZone). Timers now really fire on Advance.
• Story243's render double delays on the injected clock, so the per-unit render budget is decided by due order, never by wall-clock milliseconds. PullUnitsRacingTheRenderBudgetAsync advances one budget-vs-render pair per step.
• AdScriptWriter's per-attempt Llm:TimeoutSeconds budget is a CancellationTokenSource(delay, timeProvider) linked with the caller's token, replacing wall-clock CancelAfter. Production behaviour is unchanged; a never-answering LLM resolves to Timeout when fake time passes the budget.
• Story442 specs green: 4 pins (Architecture), 6 handoff facts (Orchestration), 2 re-ask facts (Ads).

🔁 Fifty runs (T485 wire, dev box)

for i in $(seq 50); do dotnet test tests/GenWave.Orchestration.Tests --no-build --filter "FullyQualifiedName~RendersRideThePerUnitBudget|FullyQualifiedName~FeatureHandoffBudgetOnAFakeClock" | grep -q "Passed!" || echo FAIL $i; done
# (no FAIL lines) 50/50
Passed!  - Failed:     0, Passed:     7, Skipped:     0, Total:     7, Duration: 120 ms - GenWave.Orchestration.Tests.dll (net10.0)
for i in $(seq 50); do dotnet test tests/GenWave.Ads.Tests --no-build --filter "FullyQualifiedName~FeatureAdReaskBudgetOnAFakeClock" | grep -q "Passed!" || echo FAIL $i; done
# (no FAIL lines) 50/50
Passed!  - Failed:     0, Passed:     2, Skipped:     0, Total:     2, Duration: 71 ms - GenWave.Ads.Tests.dll (net10.0)

✅ Gate

dotnet test GenWave.sln --filter "Category!=Integration" -- xUnit.MaxParallelThreads=3: all nine projects green (Orchestration 443 passed / 3 skipped, Ads 222 / 3, Host 2909 / 203, Architecture 150 / 10; 0 failed).

📝 Follow-up (out of scope)

ScenarioEachAttemptHasItsOwnTimeoutBudget in Tts.Tests' Story390 still spends about 3.6 s on the real clock; it can ride the fake clock now that the writer's budget is TimeProvider-driven.