Speed up deterministic screenshot recapture

[ariofrio]

Summary

• check screenshot relevance in a shallow, uncontainerized job and start the pinned renderer only when needed
• install the harness and plugins as one workspace, building their shared bb-app and node-pty once
• keep rendering serial and condition-driven while reporting phase timings

Measurements

Path	Original	Final	Change
No-capture required check	68s	13s	-81%
Full required check	275s	203s–231s	-16–26%
npm run screenshots	209s	125–132s	-37–40%

The final post-rebase capture reproduced all 32 committed PNGs.

Alternatives considered

• Production-only install: saved 2s in CI while making the build depend on current dependency classifications.
• Reusable compositor: saved 0.8s from total capture and introduced persistent renderer state.
• Parallel plugin builds: saved at most 1.79s; plugin installation and reload must still mutate one bb serially.
• Custom image: would still pull the 25–28s Playwright base while adding image publishing and larger layers.
• Dependency cache: an 834 MB native dependency tree is not justified against a 22–29s clean install.
• Concurrent capture: changed Linux output, so capture remains serial.

Verification

• npm test — 81 passed
• all five plugin builds passed
• the workspace lock contains one bb-app and one node-pty
• the post-rebase pinned-container capture reproduced every committed PNG