feat: make model generation budgets configurable

[minixalpha]

Replace the fixed 8192-token reply limit with a configurable generation budget, defaulting to 32768. Add --max-tokens and ANTHROPIC_MAX_TOKENS, including the existing settings-file environment configuration, with CLI > environment > settings > default precedence. Validate the effective positive integer before model execution and apply it to every reply in interactive and headless modes.

Forward budgets through Harbor and record the effective limit in startup output, Event Journals, and ATIF separately from usage. Preserve compatibility with older journals and existing truncation behavior.

Validation: 242 core tests and 22 Harbor adapter tests passed, covering precedence, request forwarding, real SDK serialization, invalid values, recorded limits, and truncation. Python 3.13 and 3.14 CI passed. The English development notes were regenerated from the complete Chinese source; examples, links, and experiment figures were checked.

Benchmark diagnostics recorded two regex-log passes at 32768, two compressor truncations at 32768, and one compressor pass at 65536 with a longer deadline. That successful run peaked at only 32073 output tokens, so it does not establish a benefit from doubling the cap. The other 18 tasks passed 5/18 at 65536, versus 8/18 in the 8192 baseline; five supplemental attempts added no passes. Different source revisions, deadlines, runtime conditions, and infrastructure faults prevent a controlled budget-only comparison. No benchmark-wide improvement is established. The default remains 32768, with 65536 available as an override.

The PR includes changelog entries, bilingual user and Harbor documentation, development notes, and structured experiment results indexing retained local trajectories. Both root READMEs are unchanged; research-note files are unchanged.