fix(compressor): keep truncated tool_call arguments as valid JSON

[handsdiff]

Summary

context_compressor.py Pass 3 truncates large tool_call arguments strings by byte-slicing to 200 chars and appending the literal sentinel ...[truncated]. The result is invalid JSON — mid-value, unterminated string.

Some providers (notably Anthropic via LiteLLM) re-parse function.arguments when rebuilding tool_use blocks. Once a session is compressed with any tool_call whose original args exceeded 500 chars (routine for patch), the next API call fails with:

HTTP 400: litellm.BadRequestError: AnthropicException -
Failed to parse tool call arguments for tool 'patch' (Anthropic tool invoke).
Error: Unterminated string starting at: line 1 column 35 (char 34).

The 3 retries are identical payloads, so they all fail identically, and every subsequent user turn replays the corrupted history. The session is permanently stuck and the user sees only a canned error fallback.

Fix

Replace the byte slice with a compact JSON sentinel that preserves provenance (original length + 200-char preview) while staying parseable:

{"_truncated": true, "_original_length": 1247, "_preview": "..."}

Observed in the wild

A production session had 11 prior patch tool_calls poisoned by this truncation after compression. Every subsequent user message triggered a new failure cycle — user saw 3 cryptic LiteLLM 400 error strings back-to-back with no way to recover (the poison is in server-side history).

Test plan

• New unit test test_pruned_tool_call_args_remain_valid_json that asserts the pruned arguments are parseable via json.loads and carry a sensible sentinel shape.
• All existing TestTokenBudgetTailProtection tests still pass.