fix: handle truncated model responses explicitly - #36

Model replies ending with max_tokens were recorded as completed. Report response_truncated, print a stderr diagnostic, skip tools from the truncated reply, and preserve partial output, usage, and cost accounting. Keep subsequent interactive requests valid by retaining text and a truncation notice without replaying unexecuted tools or unfinished thinking. Headless exit status remains 0; the verifier determines task success.

New Event Journals use schema v2, with v1 replay and projection still supported; public trajectories remain ATIF-v1.7. Update the changelog and bilingual protocol/CLI documentation, regenerate the complete English development notes from the Chinese source, and record the two-case benchmark rerun. Increased generation budgets and automatic continuation are documented as pending requirements; this implementation still stops at the unchanged response limit.

Validation:

• Core test suite: 205 passed; Harbor adapter tests: 16 passed.
• CI passed on Python 3.13 and 3.14.
• Real Docker reruns of regex-log and write-compressor: 0/2 passed at 8192 tokens per response, with no Harbor exceptions. Both runs correctly report response_truncated, and both trajectories pass the official Harbor ATIF validator. The required output files were not created before truncation.