perf: add a lightweight runtime entry point and bundle-size budgets

[vamgan]

Summary

Add a genuinely lightweight Askable entry point and enforce browser bundle-size budgets in CI.

Why

Askable's runtime benchmarks are healthy, but bundle weight is not currently budgeted. A measured consumer that imports only useAskable from @askable-ui/react/core still bundles approximately 63.5 KB minified / 18.2 KB gzip on main (40a3374). Importing from the normal React entry is approximately 65.0 KB minified / 18.5 KB gzip.

The main reason is structural: createAskableContext() is implemented by a large class in packages/core/src/context.ts, so optional source resolution, packet/protocol, async subscription, viewport, serialization, and related behavior cannot be tree-shaken at the method level. The current /core framework subpaths reduce exports but do not provide a truly small runtime.

Package evidence after build:

• @askable-ui/core: ~147 KB tarball / ~693 KB unpacked
• @askable-ui/react: ~77 KB tarball / ~389 KB unpacked
• core package source maps account for roughly 43% of built dist

Runtime performance itself is currently good: all 9 existing core benchmark budgets pass, including 10k-element observation in ~26 ms. This proposal is about transfer/parse weight and preventing future regressions.

Proposed direction

1. Add a minimal runtime entry point, for example @askable-ui/core/lite, containing only:
   • annotate/observe/select/push
   • current focus + bounded history
   • sanitization
   • basic prompt serialization
   • lifecycle/events
2. Keep advanced capabilities opt-in through explicit subpaths/plugins:
   • sources/async resolution
   • context packet protocol
   • inspector
   • region/text capture
   • extended browser sources
3. Add matching framework entry points such as @askable-ui/react/lite.
4. Add a deterministic esbuild-based bundle check to CI with committed gzip budgets.
5. Consider omitting source maps from published tarballs or making that an explicit release tradeoff while retaining declarations.

Acceptance criteria

• A documented lightweight entry point exists without breaking current imports
• A minimal React consumer is at or below an agreed target (suggested initial target: 10 KB gzip)
• CI measures representative core and adapter consumer bundles
• CI fails on budget regressions with an actionable report
• Inspector/capture/source features remain available through explicit imports
• Runtime benchmark budgets remain green