fix(responses): preserve callable identities across translation and shims

[M1k0t0]

Responses requests can reuse a callable name across namespaces, function/custom declarations, and replayed history. Translation can conflate these identities, while server-tool shims can mistake a namespaced client call for an injected helper. For example, a historical function read must remain distinct when the next request declares a custom tool named read.

This change preserves those identities through translation, continuation, and hosted-tool execution:

• Use one translator-owned mapping for declarations, additional_tools / tool_search_output, replay, and selectors in both Responses → Chat Completions and Responses → Anthropic Messages. Restore source names, namespaces, callable kinds, argument events, and echoed tool fields on the return path.
• Keep names containing . or __ literal unless an explicit namespace is present. Allocate distinct aliases for callable kinds and namespace tuples, preserve namespace descriptions, and bound collision lookup and long-name processing.
• Reserve client declarations and history when naming server-tool helpers. Dispatch only unnamespaced helper calls, reject identity changes between item creation and completion, and preserve allowed_tools subsets and original choice echoes across helper iterations.
• Keep native Responses payloads intact and isolate translated provider payloads from canonical history and response echoes. Cover both translation APIs directly and the HTTP continuation, retry, compact-expansion, and server-tool paths.

Validation on bddaf78eb:

• pnpm run verify passed, including lint, typecheck, 6,434 tests across 588 files, repository checks, and the web build.
• Installer harness: 66 passed, 44 skipped because this host lacks PowerShell or the pinned Codex 0.144.5 executable.