fix(tools): isolate OpenAI middleware clients

[abhinav7x94]

Summary

• return a descriptor-preserving OpenAI middleware facade instead of mutating the supplied client
• canonicalize facades across rewraps, duplicate module copies, and HMR so tenant middleware never stacks
• preserve original resource receivers and forward OpenAI RequestOptions
• add offline Chat, Responses, persistence, frozen-client, prototype, and duplicate-module regressions

withSupermemory previously overwrote create methods on the shared client. Wrapping the same client for tenant A and then tenant B caused a B request to retrieve and inject both tenants' memories and, with automatic persistence, write the conversation to both tenant threads.

Validation

• focused isolation suite — 5/5
• full tools suite — 95 passed, 30 skipped; only two pre-existing suite-load failures (missing live API key and bad legacy import)
• production tsdown build — passed
• Biome and diff check — passed
• changed files type-clean; package typecheck remains baseline-red and has one fewer error than main
• two independent reviews — ship

Duplicate audit covered all 99 open PRs. #1244, #1286, #1465, and #1489 touch the middleware for unrelated work; none solve client mutation/isolation. #1453 touches a distant README hunk.