feat: optimize context compression algorithm

[YYH211]

Summary

• Optimize request-message, large-content, and tool-response context compression behavior.
• Remove hard max_tokens truncation from compression LLM calls and rely on prompt budget plus final token validation.
• Add deterministic over-budget fallbacks and source anchor preservation to improve fidelity and structure.
• Update compression tests for budget validation, fallback behavior, and benchmark adapter expectations.

Related

• Requires the pro benchmark update: https://github.com/labring/fastgpt-pro/pull/978

Validation

• pnpm --filter @fastgpt/service test test/core/ai/llm/compress/index.test.ts
• pro benchmark validation:
  • pnpm typecheck
  • pnpm test

[github-actions]

Coverage Report

Status	Category	Percentage	Covered / Total
🔵	Lines	14.04%	1158 / 8246
🔵	Statements	14.01%	1213 / 8654
🔵	Functions	12.53%	248 / 1979
🔵	Branches	12.14%	550 / 4527

File Coverage

No changed files found.

Generated in workflow #567 for commit d30b99f by the Vitest Coverage Report Action

[github-actions]

✅ Build Successful - Preview fastgpt Image for this PR:

ghcr.io/labring/fastgpt-pr:fastgpt_d30b99f9d05e775001d4a1ae5cbcc7974ddb8ad9

🕒 Time: 2026-06-03 21:34:35 (UTC+8)

[github-actions]

✅ Admin Preview Image Ready!

ghcr.io/labring/fastgpt-pr:admin_d30b99f9d05e775001d4a1ae5cbcc7974ddb8ad9

🕒 Time: 2026-06-03 21:37:41 (UTC+8)