fix(offload): invalidate token cache after stripping deleted tool_use blocks

[akhilesharora]

Description | 描述

context-token-tracker.ts caches per-message token counts in a WeakMap keyed on message identity plus _offloaded. Two offload hooks strip deleted tool_use blocks out of mixed assistant messages in place with content.splice(...) but never call invalidateTokenCache(msg) and never set _offloaded:

• before-prompt-build.ts (splice at line 109)
• llm-input-l3.ts (splice at line 1377)

So the next buildTiktokenContextSnapshot returns the stale pre-strip count, and since the strip runs just before the guard and emergency snapshots, the inflated total can cross the thresholds and over-delete real messages. Every other in-place mutation already invalidates the cache (l3-helpers.ts:219/257/302, llm-input-l3.ts:1019); these two were the only ones that skipped it. This calls invalidateTokenCache(msg) after each strip loop, only when a block was actually removed, and adds the import to before-prompt-build.ts.

Related Issue | 关联 Issue

Closes #233

Change Type | 修改类型

• Bug fix | Bug 修复
• New feature | 新功能
• Documentation update | 文档更新
• Code optimization | 代码优化

Self-test Checklist | 自测清单

• Verified locally | 本地验证通过
• No existing features affected | 无影响现有功能

Additional Notes | 其他说明

Reproduced against the real context-token-tracker.ts: seed the cache, strip a deleted tool_use block in place, snapshot again.

Before:

after strip, no inval : 537
ground truth (fresh)  : 18
drift                 : 519

After (invalidate following the splice): 18, drift 0.

npm test -> 47 passed. No new test added: the strip loops live in non-exported hook closures, so a unit test cannot drive them without a mocking harness the repo does not have. The runtime repro above is the evidence.

[Maxwell-Code07]

@akhilesharora Clean fix! Adding invalidateTokenCache() at both splice sites is the right approach — targeted and thorough with regression tests. Nice work 👍