Accept sanitised Kimi tool-call ids and return the canonical form

[Bdwg222]

Motivation

Fixes #2281 (part 1 of 2).
_parse_kimi_tool_calls requires the id immediately before <|tool_call_argument_begin|> to match functions.<name>:<n>. Kimi K2 models also emit functions_<name>_<n> — they do so whenever a client has echoed a sanitised id back in the conversation history, which is common (OpenAI-style clients restrict ids to [A-Za-z0-9_-]). That produces a hard parse failure, surfaced to the client as an ErrorChunk whose message is the raw tool-call markup (streaming) or a bare 500 (non-streaming). In practice the second tool call in any agent session fails deterministically.

Changes

• src/exo/worker/engines/mlx/utils_mlx.py — _parse_kimi_tool_calls: the id regex now accepts . or _ after functions, and : or _ before the index. The returned id is always the canonical functions.<name>:<n> regardless of which form the model emitted.
• src/exo/worker/tests/unittests/test_runner/test_kimi_tool_parser.py (new) — unit tests for the canonical form, the sanitised form, underscores inside function names, a missing functions. prefix, multiple calls with mixed forms, and the still-rejected no-index case.

Why It Works

The trailing [:_]\d+ is unambiguous even when the function name contains underscores (functions_read_file_3 → read_file, 3): the lazy name group can only stop where separator + digits + the argument marker follow, so it cannot split the name early.
Returning the canonical id also breaks the feedback loop that causes the bug — the client now receives functions.<name>:<n>, echoes that back in the history, and the model keeps emitting the form it was trained on.

Test Plan

Manual Testing

Hardware: two Mac Studio M3 Ultra 512GB, two-node MLX ring, serving moonshotai/Kimi-K2.7-Code.

• Reproduced the failure with the two-turn request in Kimi K2 tool-call parsing fails when the model emits functions_<name>_<n> ids; failure surfaces as an error chunk containing the raw tool-call markup #2281 (sanitised id in history) → 500 / error chunk.
• Verified the patched parser against the exact failing payload: functions_Write_3 → functions.Write:3.
• Multi-step agent run via Kimi Code CLI (LiteLLM proxy → Exo), ~12 consecutive tool calls including Read, Bash, Write, git and gh pr create; previously failed on the second call, now completes.

Automated Testing

• New test_kimi_tool_parser.py covers the cases listed under Changes.
• Existing test_parse_tool_calls.py unaffected; behaviour on parse failure is unchanged in this PR.

Not in this PR

Part 2 of #2281 — having parse_tool_calls degrade to plain text on a parse failure instead of emitting finish_reason="error" — is a deliberate behaviour change (the current behaviour is asserted in test_failed_parse_yields_text), so it is kept out of this PR for discussion on the issue first.