feat(sight): agent crash full-stack observability

[jfeng18]

Summary

Agent 进程层级全栈可观测：血缘树 + 精准 crash 检测 + blast radius 量化 + token 按层级归因 + CLI crash report + token --by-type 聚合查询。

agent 崩了，系统知道崩没崩、为什么崩、影响多大、崩的时候在干什么、哪个子进程在烧钱。

Strategic positioning

This PR is the AgentSight foundation for the OS-middle-layer recoverability and task-level observability substrate: process supervision, exit-reason capture, Agent/SubAgent/Tool lineage, crash blast-radius reporting, and token attribution by process type.

It deliberately does not implement the scheduling/control-plane pieces from the broader Agent Infra roadmap: no cgroup QoS control, no LSM enforcement, no transparent MaaS cache proxy, and no closed-loop task recovery orchestration. Those can build on this observability substrate later.

Problem

1. crash 检测靠猜：用 pending LLM calls 判断 crash，SIGSEGV 在两次调用间 → 漏判，正常退出 SSE 没收完 → 误判
2. 子进程完全不可见：SubAgent/Tool 崩溃静默丢弃（scanner 只认 root agent）
3. token 无层级归因：不知道哪个子 agent 在烧钱，无法按 Agent/SubAgent/Tool 分层分析
4. 启动时已有 agent 不在 lineage tree：scanner 初始扫描发现的 agent 没有 lineage 节点

Solution

层	改动	效果
BPF 采集	procmon.h 加 exit_code，BPF_CORE_READ task_struct	crash/normal 从内核事实判定
血缘树	lineage/mod.rs：pid→ppid 树 + Agent/SubAgent/Tool 分类 + root_agent_ancestor	进程层级结构化
crash 判定	classify_exit() 纯函数替代 pending-calls 猜测	假阴性/假阳性修复
blast radius	ProcessType → total_session_loss/partial/recoverable	量化崩溃影响，severity 按类型分级
SubAgent/Tool crash	Exit handler 扩展：lineage 分类的子进程 signal crash 也检测	子进程不再静默
token 归因	LLMCall/PendingCallInfo 加 process_type，SQLite+Logtail 全链路	按层级分析消耗
token --by-type	GenAISqliteStore 聚合查询 + CLI agentsight token --by-type	一行看哪层在烧钱
启动补种	构造函数末尾 insert_and_classify existing_agents	已有 agent 也在 lineage tree
CLI	interruption get 结构化 crash report（信号/进程树/blast radius/process type）	一眼看懂崩溃现场
bug fix	pid_agent_name_cache PID 复用清理 + proctrace exit_code TODO 修复	数据准确性
race fix	proctrace Exit 不再删 lineage node，defer 到 procmon Exit handler	防止 crash 事件因事件排序竞态静默丢失
CI fix	background.rs tmp_dir 用 CARGO_MANIFEST_DIR 替代 temp_dir	修复 llvm-cov 下 SQLITE_READONLY

Crash detection truth table

场景	exit_code	旧判定	新判定
SIGSEGV 无 pending	11	漏判	crash(Critical, blast=total_session_loss)
SIGKILL (OOM)	9	crash(仅有pending)	crash(Critical)+oom
SIGTERM 优雅停止	15	crash(若有pending)	normal(graceful)
exit(0)+pending SSE	0	误判(crash)	no crash event (pending lifecycle remains separate)
Tool SIGSEGV	11	静默丢弃	crash(Medium, blast=recoverable)
SubAgent SIGSEGV	11	静默丢弃	crash(High, blast=partial)
子进程 exit(1) 跟随父死	256	N/A	不产出(非 signal，正常清理)

CLI examples

$ agentsight interruption get <id>

  --- Crash Report ---
  Cause:        SIGSEGV (signal 11, coredump: true)
  OOM Killed:   no
  Process Type: agent
  Blast Radius: total_session_loss
  Children at Exit: [5678]

  Process Tree at Crash:
    `-- cosh [1234] (agent)
        `-- node [5678] (tool)

$ agentsight token --by-type --hours 24

Token Usage by Process Type (last 24h)

TYPE           CALLS        INPUT       OUTPUT        TOTAL
------------------------------------------------------------
agent              15       12,345        8,901       21,246
sub_agent           3        2,100        1,500        3,600
tool                1          200          100          300
------------------------------------------------------------
TOTAL                                                25,146

Verification

Check	Status
cargo fmt/clippy/test (1031 tests)	pass
check-arch-boundaries.py	pass
BPF verifier load (kernel 6.6.102+)	E2E pass
SIGSEGV root agent → crash + blast_radius=total_session_loss	E2E pass
SIGTERM → no crash event	E2E pass
子进程 exit(1) → no crash (噪音修复)	E2E pass
process_type in genai_events SQLite	E2E pass
CLI crash report 含 Process Type + Blast Radius	E2E pass
token_usage_by_process_type 排除 pending 行	区分性测试
Adversarial workflow review (2 rounds, 14 agents)	0 critical/high

What is NOT changed

• Not a cgroup/QoS scheduler: no cpu.weight, cpu.max, burst control, or dynamic resource enforcement
• Not an LSM/sandbox policy engine: no AppArmor/SELinux/eBPF-LSM allow/deny control
• Not a transparent MaaS cache proxy: no socket redirection or LLM response cache interception
• Not closed-loop recovery orchestration: crash reports/checkpoint context only; resume/retry policy remains out of scope
• health/checker.rs：serve 进程没 BPF exit_code，保持 pending-calls 逻辑（已知限制）
• procmon pid namespace：保留 host tgid，容器 PID namespace 场景标注限制

[CLAassistant]

All committers have signed the CLA.