Optimize TypeScript SDK VM memory

[w0nche0l]

Summary

• add fresh-process bundle/import/scenario characterization with raw, gzip, and Brotli sizes from minified tree-shaken builds
• add deterministic SSE fragmentation, error, completion, and cancellation coverage
• split codec-only base64 helpers from Zod adapters using Speakeasy persistent edits
• isolate the broad Responses request/event schema graphs across 20 fresh runs
• record rejected index-free and dynamic-error prototypes so they are not retried without changed prerequisites

Twenty-run attribution

Median retained import heap:

• OpenRouterCore: 965,552 bytes
• responsesSend: 41,988,580 bytes
• Responses request schema: 21,757,296 bytes
• all stream-event schema: 21,580,420 bytes
• text-delta-specific schema: 1,087,232 bytes
• completed-event schema: 18,010,468 bytes

A text-delta-specific schema is ~95% lighter than the all-event union. Lazy event dispatch alone only shifts the peak because completion eventually loads an 18 MB graph. A material reduction needs generator-owned operation-private, discriminator-aware validators.

Rejected prototypes

• useIndexModules: false: pinned generation creates naming/type collisions and would require unsafe aliases/public renames.
• status-specific dynamic error imports: with the current non-splitting bundle, median sender import regressed 41,988,580 → 99,969,304 bytes and bundle size 213,200 → 357,667 bytes. Fully reverted.

Safe retained optimization

The codec split reduces the transport bundle 118,279 → 118,098 raw bytes and median retained transport import heap 973,592 → 965,016 bytes.

Verification

• all characterization bundles are minified with explicit tree shaking
• format, ESLint, typecheck, build, and 207 unit tests pass
• 25 live OpenRouter E2E tests pass with OPENROUTER_TEST_KEY
• frozen lockfile install and pinned Speakeasy regeneration pass

Linear: https://linear.app/openrouter/issue/ECO-2751
Parent: https://linear.app/openrouter/issue/ECO-2747