feat(mind): add PageIndex-style structure retrieval

[keli-wen]

Summary

Adds PageIndex-style structure retrieval as a self-contained artifact plus a pure-agentic reasoning retriever, and refreshes the design principles that shape both. After review the original "vectorless empty shell + library page-refill" design was replaced by a self-contained one, so the final shape is:

• knowledge: a shared StructureTree base and a source-bound PaperStructureTree whose leaf nodes carry their own page-cited text (internal nodes stay navigational). from_draft mints all identity, links, and citations, and copies minimal provenance (ArtifactMeta: as_of + a light source ref + source_content_hash) that is kept out of id / content_hash, so identity stays reproducible and the tree is storable standalone.
• flows: PaperFlow(cfg) is a config-bound flow — bind the cfg once, build(input) takes only the operand, and the cfg type selects the knowledge shape (PaperStructureCfg → PaperStructureTree). paper_flow remains the compatibility function for the semantic chunk/summary shape. A pipeline is pure processing and binds no library.
• library: stores a self-contained tree standalone via put(tree) / open_structure(id) (no source or chunk set required); SQLite schema v5 drops the artifact→source foreign key with a migration; resolve returns a node's stored content with no query-time refill.
• mind: AgenticRetriever(RetrievalCfg) is pure-agentic — an LLM agent reasons over one self-contained structure and returns evidence values (content included; locator is optional provenance), with no library binding. Mechanical retrieval (semantic / BM25) stays in library / rag; a future hybrid path composes the two.
• preprocess: deterministic outline signals (headings, table-of-contents pages, printed-to-physical page offset) feed the single bounded structuring draft call.
• defaults / robustness: PaperStructureCfg and RetrievalCfg default to gpt-5.6-luna; AgenticRetriever wraps the SDK MaxTurnsExceeded in RetrievalError and raises RetrievalCfg.max_turns to 20 so agentic traversal has head-room.
• design principles: contexts/design/mind/retrieval.md, operations/orchestration.md, operations/naming.md, AGENTS.md, CLAUDE.md, and the quantmind-dev skill now state: bind config (not the operand) for batch reproducibility; the cfg type selects shape/strategy (typed dispatch, not a class hierarchy); artifacts are self-contained (content + provenance); persistence and retrieval are downstream and retrieval returns values; and mind is the reasoning layer while rag / library are the mechanical retrieval plane.
• observability: scripts/verify_structure_e2e.py is a live smoke that builds, persists, reopens, and retrieves under the default model, wired as a path-filtered, non-required structure job in .github/workflows/e2e.yml.

Related Issue

Part of #95; design context in #122. Live-smoke follow-ups filed: #126 (structured output fails on json_object-only providers) and #127 (surface token-usage / cost).

Verification

• bash scripts/verify.sh — pass: Ruff format + lint, basedpyright (0 errors), import-linter contracts, and 382 tests at 84.35% coverage.
• Live structure smoke: python scripts/verify_structure_e2e.py — PASS with gpt-5.6-luna (self-contained tree covering every page, put / open_structure round-trip identical, agentic retrieve returns evidence).

Checklist

• The title uses English Conventional Commit format: type(scope): summary.
• The related issue or design discussion is linked when applicable.
• bash scripts/verify.sh passes.
• Every applicable live-network component smoke test passes, or this PR states why none applies.
• Public behavior has focused tests, an example, and documentation where applicable.
• The PR is complete, small, and contains no unrelated changes.