Lazy-load Light and Sun stylesheets

[elkimek]

What changed

• defer the seven Light & Sun stylesheets until a visible Light UI entry point is opened
• keep background Light/Sun session analysis on the existing module-only loader so it does not fetch presentation assets
• preserve the original cascade position with an ordered stylesheet anchor and make concurrent loading single-flight and retry-safe
• move the two cross-feature style consumers (marker history controls and context-card Ott badges) into their eager owning bundles
• retain all deferred resources in the service-worker app shell for offline first use
• add browser coverage for cold startup, ordering, concurrency, failure/retry, dashboard widgets, sidebar assessment, route rendering, and Firefox offline navigation
• make the Light Today browser flow wait for the observable lazy-loaded UI state instead of a fixed runner-dependent delay

Why

Returning users paid for seven Light-only stylesheets on every startup even when they never opened Light & Sun. This moves that cost to first visual use without delaying background analysis or weakening offline behavior.

Impact

Measured on the same fresh mobile returning-user scenario with browser cache and service workers disabled:

• requests: 448 → 441 (-7)
• compressed transfer: 1,900,879 → 1,857,409 bytes (-43,470)
• decoded bytes: 5,843,728 → 5,650,563 (-193,165)

The ratcheted cold-load budget and architecture notes are updated, and the app version advances to 1.10.356.

Validation

• COVERAGE=1 PORT=8265 ./run-tests.sh
  • 487 Vitest tests passed
  • 396 Chromium tests passed
  • 472 responsive/theme checks passed
  • combined function coverage passed at 67.11% against the 67.10% ratchet
  • Firefox smoke and offline first-use coverage passed in the preceding full run
• Light Today browser fixture passed 10/10 consecutive stress runs
• real sidebar-to-lazy-Light assessment integration passed 5/5 consecutive stress runs
• npm run quality — 11/11 guardrails passed
• TypeScript typecheck, JS check, architecture, vendor integrity, and git diff --check passed
• cold-load measurement reproduced 441 requests / 1,857,409 transfer bytes / 5,650,563 decoded bytes