fix: bound retained payloads and avoid stream cloning

[satvikpadhiyar]

Summary

• Cap persisted request and response bodies before writing request_payloads, while keeping uncapped payloads for linkage and usage extraction.
• Make CF_MAX_PAYLOAD_BYTES configurable and document it.
• Skip startup SQLite compact/checkpoint work that can spike memory on large existing databases.
• Track streaming response chunks inline instead of cloning the upstream Response, avoiding extra Bun stream branches for long Claude/Anthropic streams.

Why

• Large Claude Code request/response payloads can accumulate in SQLite and make startup maintenance load or rewrite huge data.
• Response.clone() on streaming upstream responses creates an extra reader branch for analytics. In Bun this can add memory pressure and was observed locally around ECONNRESET/malformed-response failures.

Verification

• bun test packages/proxy/src/response-handler.test.ts packages/proxy/src/post-processor.worker.test.ts packages/database/src/database-operations.test.ts packages/proxy/src/usage-worker.test.ts
• bun run typecheck
• bunx --bun biome check packages/database/src/database-operations.ts packages/proxy/src/post-processor.worker.ts packages/proxy/src/post-processor.worker.test.ts packages/proxy/src/response-handler.ts packages/proxy/src/response-handler.test.ts packages/proxy/src/stream-tee.ts packages/runtime-server/src/startup-maintenance.ts docs/configuration.md

Notes

• Full repo lint was not used as the gating command because the existing tree has unrelated formatting failures outside this PR scope.