fix: namespace repeated tool call IDs

[ting-hong-shieh]

What

• Generate occurrence-aware Anthropic tool_use IDs from the source response ID and tool index.
• Restore the provider's original tool-call ID when Anthropic tool calls and results are replayed to an OpenAI-compatible backend.
• Cover buffered and streaming translation, including fragmented stream fields and missing-ID fallback behavior.

Why

OpenAI-compatible providers may reuse IDs such as call_0 in consecutive responses. Passing those IDs through unchanged creates duplicate Anthropic tool_use IDs, which can mispair tool results and leave agent workflows looping.

Addresses NVIDIA-NeMo#502.

Depends on NVIDIA-NeMo#397 and is intentionally stacked on fix/reversible-tool-use-ids.

Before / after

Before:

chatcmpl-first  / call_0 -> Anthropic call_0
chatcmpl-second / call_0 -> Anthropic call_0

The repeated Anthropic ID made replay ambiguous.

After:

chatcmpl-first  / call_0 -> Anthropic sy64o_<first occurrence>
chatcmpl-second / call_0 -> Anthropic sy64o_<second occurrence>

The Anthropic-facing IDs are distinct, while both replay to the provider's original call_0.

How tested

• uv run ruff check .
• uv run mypy switchyard
• uv run pytest tests/ -v — 147 passed, 2 skipped
• cargo fmt --all --check
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test --workspace
• Buffered and streaming regression tests replay repeated call_0 IDs correctly

Checklist

• No new classes or public symbols.
• Unit/regression tests added for the bug fix.
• No customer-facing documentation change required.
• Commit signed off per the DCO.

Notes for reviewers

• This draft should remain stacked on fix(translation): preserve tool IDs across Anthropic NVIDIA-NeMo/Switchyard#397 until that PR merges.
• Occurrence uniqueness follows the issue's suggested invariant and assumes source response IDs are unique across responses.
• No live provider workflow was run; the multi-turn ID pairing and replay path are covered deterministically by translation tests.