feat(cph): implement nested outline manifest and batch/combined export

ADR-0029 — nested outline manifest, supersedes ADR-0008's flat [[parts]]:
- cph-model: recursive loader over manifest.toml containers / element.toml
  leaves; Lesson.parts (pure elements, DFS order) + Lesson.outline (elements
  interleaved with section headings at their DFS-open position); rejects
  ambiguous/incomplete folders and root-vs-container table misplacement
- cph-diag: new DiagCode::ManifestMalformed for carrier-document structure
  errors (discharges an existing TODO)
- cph-typst: augmented manifest now serializes the outline (element/section
  entries) instead of a flat parts array
- render/lib.typ: render-lesson renders section headings at their depth
- examples/TH-141 migrated to 5 nested section containers + 3 root segments,
  byte-identical element order; smoke-verified via cph check/build + pdftotext

ADR-0030 — batch & combined export, extends ADR-0009/0011:
- cph build with no --target batches every declared target (repeatable
  --target for an explicit subset); any target failure => non-zero exit,
  per-target ledger, independent per-target execution
- cph-model: bundle.toml loader (directory + [info]/[targets.*]/ordered
  lessons with per-lesson target overrides)
- cph-typst: augmented bundle manifest (path-prefixed member outlines),
  Engine::{compile_check_bundle,build_bundle_pdf}
- render/lib.typ: render-bundle assembles member lessons under per-lesson
  headings, depth-shifts their own section headings, resets example/lemma
  counters at each lesson boundary by default
- cph-cli: `cph bundle <path> --target <name>` subcommand, same batching
  contract as `cph build`
- new bundle fixtures/tests (cph-model unit + cph-typst through-template PDF
  compile), smoke-verified via a real 2-lesson merged PDF

Verification: cargo fmt/clippy/test clean across the workspace (68 tests);
real cph check/build/bundle runs against TH-141 and a bundle fixture, PDF
content inspected via pdftotext.

examples/TH-141/收束/segments/各模型对水的预测对照/textbook.typ

@@ -0,0 +1,16 @@
+把本章四种模型对常温水液气界面 $sigma_(L G)$ 的预测整理如下：
+
+#figure(
+  table(
+    columns: (auto, auto, auto),
+    align: (left, left, left),
+    table.header[*模型*][*$sigma$ 预测（$"N/m"$）*][*相对实测 $0.072$*],
+    [量纲分析], [$tilde.op 0.1$], [量级正确],
+    [Stefan ($zeta = 1\/2$)], [$approx 0.38$], [偏大约 5 倍],
+    [缺键模型 ($zeta = 3\/4$，FCC (111))], [$approx 0.13$], [偏大约 2 倍],
+    [Eötvös 规则（外推）], [$approx 0.072$], [量级与具体值都接近],
+  ),
+  caption: [本章各模型对常温水 $sigma_(L G)$ 的预测]
+) <模型对照表>
+
+从表上读出的事实有两条：其一，所有模型都能给出对的量级；其二，量纲分析与 Stefan 这类"几乎不假设"的模型反而偏离最大，缺键模型代入具体晶面 $zeta$ 后精度提升一档，而完全唯象的 Eötvös 规则反而最接近实测。最简单的微观模型并不是最准的——粗略的微观模型给出量级，唯象的拟合规则给出具体值。