fix: make preview startup and task saves resilient

[nquyencubas523-cloud]

Summary

• forward valid small startup IDR packets instead of reopening the preview encoder and waiting indefinitely for a larger keyframe
• route single and batch task saves through one service-layer SaveOrUpdateTask operation
• validate schedules and resource admission before creating tasks or changing parameters, areas, schedules, or enable state
• restore the previous parameters and areas when a parameter update fails
• only apply resource admission to real disabled-to-enabled transitions

Root cause

Preview startup treated a valid keyframe as invalid based only on encoded packet size. The encoder was reopened repeatedly, so channel 1 algorithm overlay preview could fail until the device was restarted.

Task saving performed parameter, area, and strategy updates before SwitchTask(true) checked the resource score. When that final check returned ResourceLimit, the API reported failure after task state had already been persisted, with no transaction or compensation path. Batch apply used the same sequence.

Impact

Valid preview startup frames are now published normally. A failed task save is rejected before mutation when its schedule or start-resource admission is invalid, so callers no longer receive an error after partial state has been applied. Existing enabled tasks can still update parameters while the device is busy because no new start transition is required.

Validation

• x86 build and full test suite passed: https://github.com/cosmo-wander-ai/cosmo-edge/actions/runs/30814785785
• Sophon package and test binary build/upload passed: https://github.com/cosmo-wander-ai/cosmo-edge/actions/runs/30814785449
• Sophon target tests passed directly on the test device:
  • [save-or-update]: 2 cases, 72 assertions
  • [video-task-handler]: 34 cases, 96 assertions
• resource score 200 regression test returns ResourceLimit without creating or changing a task
• controlled invalid-schedule save left taskList.json, param.json, and area.json byte-identical
• successful save and restore round trip passed on 192.168.100.1
• algorithm overlay preview returned FLV data with a valid header, about 164 ms first-frame latency, and zero preview failures
• deployed Sophon binary remained healthy with no crashes or service restarts

The Sophon CI hardware-test job is still queued on the self-hosted runner; the same generated test binary has already passed the targeted suites directly on the target device.