feat(ai-adapters,agentic): recover Write tool args truncated by max_t…

[nonoqing]

…okens

When the model hits its max_tokens limit while streaming a Write tool call, the accumulated arguments end inside an unterminated string and fail JSON parsing. Previously this surfaced as "Arguments are invalid JSON" and the deep-research agent would retry from scratch, hitting the same cap again.

• Add repair_truncated_json in the tool-call accumulator: closes the open string and any unclosed {/[ brackets. Refuses to fabricate values when truncated mid-pair (trailing ',' or ':').
• Gate recovery by tool name (Write/file_write/write_notebook). Bash, Edit, Task etc. still fail loudly — executing a partial shell command or a partial old_string is unsafe.
• Plumb a recovered_from_truncation flag through FinalizedToolCall -> ToolCall (#[serde(default)] for back-compat) -> tool pipeline.
• When recovery succeeds, prepend a warning to result_for_assistant instructing the agent to continue with a follow-up call rather than rewriting the whole file.
• When recovery isn't possible for an unsafe tool, surface a clearer error message naming max_tokens instead of generic "invalid JSON".
• 8 new accumulator tests cover the Write recovery path, the Bash refusal path, multibyte content, nested-bracket repair, and the refuse-mid-pair guards. All 23 existing tests still pass.