feat(serve): 调用链路面板展开/折叠 + has_more 截断标记 (#152)

[c2j]

概要

Closes #152

serve 模式详情面板的 Callers/Callees 调用链树此前一次性平铺渲染全部层级，且预算截断产生的"假叶子"（children: []）与真叶子在 JSON 中无法区分，复杂链路不可聚焦、截断状态误导。本 PR 实现 issue 的必做 A（折叠交互）与必做 B（has_more 标记），懒加载真分页（可选 D）按 issue 建议暂缓。

后端（507a5dd）

• TreeNode 新增 has_more: bool + more_count: usize（未展示直接孩子数，精确值）
• build_tree_dfs 返回签名改为 (Vec<TreeNode>, usize)：邻居收集（含 ancestors/builtin 过滤）提前至 early-return 前，预算耗尽、深度钳制、循环 break 三种截断统一折算为 unexpanded，push 时 has_more = unexpanded > 0 — 闭环覆盖且不误报（真叶子/全祖先环均 effective=0）
• 顶层 truncated / caller_count / trace_chain 签名保持兼容
• HTTP（tree_nodes_to_json）与 MCP（同名副本）同步输出新字段
• TDD：6 个新测试（Red → Green），构造点唯一，TUI/CLI 消费方零影响

前端（c3d7319）

• +/− 逐层折叠，默认展开 1 层（codeweb-tree-expand-depth 持久化）
• expand all / collapse all（清空显式状态 + 调整默认深度，免枚举）
• 折叠状态键 = currentTraceKey|pathKey 按目标作用域 — 同名节点跨分支不误伤，跨目标不泄漏（code review 修正项）
• has_more 节点渲染「▾ N more」注解徽章（无连接符样式），点击复用现有 navigateTo 重新获取 500 节点预算深入查看
• 折叠/展开操作零 fetch（缓存 trace 重渲染 + 保留滚动位置）；导航重置顶部
• localStorage JSON 读取加 try/catch 守卫；trace 请求 depth=50→10 消除被服务端钳制的误导

测试

• cargo test traverse — 6 个新截断行为测试 + 既有 14 个全过
• cargo test --features serve / cargo test --features full（跳过 CI 已知环境项）全绿
• cargo build --features full + clippy --features full -- -D warnings + fmt --check 干净
• 浏览器手工验收（serve + demo 图）：默认折叠/逐节点 toggle 不误触导航/expand·collapse all 及刷新持久化/徽章渲染与深入跳转/跨目标状态隔离/损坏 localStorage 容错

说明

• assets/app.js 为实际 UI 唯一真相（README 提及的 Cytoscape 画布未落地，与本需求正交）
• out of scope：懒加载真分页（现有 nodes/:id/callers|callees 分页端点已具备，待实测超限频率再立项）、CLI/TUI 输出的 more 提示

[c2j]

Review — Overall: Approve with minor suggestions

I traced the backend counting logic through all four truncation paths (budget exhausted at a child's recursion entry / budget breaking a sibling loop / depth clamp / genuine leaf) and the has_more/more_count attribution is correct at every level: no double counting, no false positives, and the two filters (ancestors + skip_builtins) are correctly applied before effective is computed so more_count is an exact value, not an upper bound. The six new tests are meaningful and map 1:1 to real behaviors (the budget-exhaustion-at-recursion-entry case is the hardest one and it is precisely locked). The TreeNode struct has a single construction site (traverse.rs:161 — verified by grep), so the additive fields are zero-breakage, and both serializers (HTTP handlers.rs, MCP tools.rs) are updated in sync.

Frontend: the collapse key design (currentTraceKey | section.path-idx) correctly isolates state per trace target AND per tree position, which natively avoids the "duplicate node across branches" trap from issue #152. Toggling is zero-fetch (cached currentDetailData + local re-render with scroll preservation), and the delegated click ordering ensures [+] never triggers navigation. The depth=50→10 change is honest (server was silently clamping anyway). localStorage guards are robust.

Nit 1 (recommend fixing before merge) — expand/collapse-all "default depth" leaks across targets and persists

expandAllTree sets treeExpandDepth = 99 and collapseAllTree sets it to 0, and neither is reset in navigateTo/hideDetail, and both are written to localStorage.

Consequence: after clicking "collapse all" once on node A, opening node B renders B's entire tree at default depth 0 — the top-level callers/callees are all tucked behind [+] and effectively invisible. It survives page reloads too. This contradicts the PR's own "cross-target isolation" goal: the explicit collapse map (treeCollapsed) is correctly scoped per target, but the default depth leaks globally.

Suggested fix: reset treeExpandDepth to 1 on navigateTo/hideDetail (or treat expand/collapse-all as session-only and stop persisting it), so expand/collapse-all acts only on the current target.

Nit 2 (cosmetic, cheap) — leaf placeholder width vs [+]/[-] mismatch

The leaf placeholder is   (~1 monospace char + trailing space) while [+]/[-] is 3 chars. For siblings at the same depth where one is expandable and one is a leaf, the node-tag/key columns shift by ~2 chars and rows look ragged. A fixed min-width (e.g. 2ch) + centering on .tree-toggle, or an equal-width placeholder, would keep the columns aligned.

Nit 3 (design, acceptable) — has_more badge hidden while node is collapsed

The if (isCollapsed) continue; skips both children and the "N more" badge, so a truncated fake-leaf at the default collapse depth only shows [+]; the user must expand before seeing "N more — click to open". Acceptable because has_more already makes the node expandable (the [+] itself signals "more below"), but if you want stronger affordance you could always render the badge for nodes with children: [] and has_more: true, independent of collapse state.

Notes (no action needed, confirmed)

• caller_count/callee_count still report the visible top-level count (smaller than reality under truncation) — pre-existing behavior, not claimed to be fixed here; the global truncated banner still covers it. Could be a follow-up issue.
• Frontend has no JS test harness and CI skips test_serve_*; manual browser acceptance is consistent with current repo state — acceptable.
• README Cytoscape mismatch correctly declared out of scope in the PR body.

Overall this is a solid, well-tested implementation of issue #152. Happy to approve once Nit 1 is addressed or confirmed as an intentional global preference.

[c2j]

感谢细致的 review，四处截断路径的追踪确认很有帮助。三个 nit 已在 92e40bd 全部处理：

Nit 1 — 已修复（采纳方案变体）：expandAllTree/collapseAllTree 不再修改全局 treeExpandDepth（该全局键已彻底移除，默认深度回归常量 TREE_DEFAULT_DEPTH = 1），改为 collectExpandableKeys() 枚举当前树的可展开节点、逐个写入按目标作用域的 treeCollapsed 键（target|pathKey）。这样 expand/collapse-all 严格作用于当前目标、经 localStorage 持久化、跨目标零泄漏，也与「导航后状态保持」验收项一致。浏览器回归确认：A 上 collapse all → B 按默认深度渲染 → 回 A 全折叠恢复；干净目标无显式键时严格遵循 depth 规则。

Nit 2 — 已修复：.tree-toggle 加 min-width:3ch; text-align:center，同层叶子与可展开节点列对齐。

Nit 3 — 已采纳增强：children: [] && has_more: true 的假叶子折叠后仍渲染「▾ N more」徽章（[+] 与徽章并存，折叠态保留深入入口）；有 children 的节点折叠后徽章保持隐藏，原行为不回归。

Notes 部分无异议：caller_count 可见性语义留作后续 issue。