fix(v2): stop compaction usage from becoming phantom system overhead (#421)

[ranxianglei]

Closes #421

Problem

V2 native compaction projects the checkpoint as an assistant carrying the compaction request's token usage verbatim (lib/v2/projection/shared.ts makeAssistantInfo). cacheSystemPromptTokens used that assistant as its calibration anchor, permanently storing the request size as "system + tool schemas" overhead. Repro from the issue: input 653137 → stored 653136 phantom; on a 400000-token model the hard-guard budget becomes −285904, producing negative-budget log spam and unnecessary tool clearing while real provider usage stays far smaller.

Fix

• lib/token-utils.ts — new calibrateSystemOverhead(state, messages): never anchors on info.summary === true assistants or assistants created before state.lastCompaction; subtracts the full wire-visible pre-anchor prefix with BPE-exact counts (excluding host-ignored user parts and ACP-owned notices). estimateSystemPromptTokens(messages, lastCompaction = 0) delegates to it.
• lib/ui/utils.ts — cacheSystemPromptTokens(state, messages, measuredSystemTokens?) merges the heuristic residual with a measured current-wire floor via max(), recording provenance in new transient field SessionState.systemPromptTokensSource (heuristic | measured). The [FIX estimateSystemPromptTokens inconsistent between nudge and acp_status after compression #255] stable-cache early return is preserved.
• lib/v2/context.ts — measures the actual outgoing system parts (Σ countTokens(event.system[].text) + rendered ACP system prompt) each transaction and passes it as the floor. V1 callers omit it → pure heuristic, unchanged.
• Invalidation — resetOnCompaction (lib/state/utils.ts) and the mid-session model-switch block (lib/messages/transform.ts) now clear the cached estimate, so first-request-after-compaction and after reload recalibrate cleanly instead of inheriting a stale/phantom value.
• lib/compress/status.ts — /acp context breakdown labels the system segment [measured] / [estimated], distinguishing actual wire accounting from heuristics (issue acceptance criterion).

No persisted-state format change (systemPromptTokens* were already transient), no config change, all API extensions are optional trailing parameters.

Verification

• New regression suite tests/v2-compaction-system-overhead.test.ts — 14/14 pass: the exact V2: compaction usage becomes phantom system overhead and negative context budgets #421 repro (fresh state + reload both store no phantom), post-compaction anchor calibrates to real wire overhead with nonnegative budget at the reported 400k limit, measured-floor/provenance behavior, invalidation paths, and estimateSystemPromptTokens inconsistent between nudge and acp_status after compression #255 compatibility (plain-text sessions calibrate byte-for-byte as before).
• Bug-catch verified: with lib/ reverted to pre-fix code, the original cacheSystemPromptTokens on the repro input stores exactly 653136; fixed code stores undefined.
• Full suite: 1424/1425 (sole failure = tests/soft-block.test.ts module-load EACCES /tmp — sandbox mounts /tmp read-only, pre-existing environmental issue, unrelated).
• npm run typecheck clean, npm run build green, touched files prettier-clean.
• Devlog: devlog/2026-09-17_v2-compaction-phantom-system/{REQ,DESIGN,WORKLOG}.md.

中文摘要:修复了 V2 原生压缩后校准锚点误用 compaction 请求用量、把 653136 tokens 当成永久 system 开销导致硬护栏预算为负(-285904)的问题——排除 summary/过期锚点、按当前实际出站 system 部件做测量下限、在压缩与模型切换时失效缓存,并在 /acp context 中标注 measured/estimated 来源。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 收到，开始核验这个 PR（不重做，按 [PR] 流程：拉分支核对代码 + 确认 CI + 报告结论）。已确认 PR 基于 2026-09-17_v2-base（V2 迁移分支），head 为单提交 64f31509，GitHub 侧 mergeable/clean。正在本地跑新回归套件和全量测试，稍后给出完整结论。