Add outline-guided retrieval path for governance analysis

[autholykos]

Context

The useful PageIndex lesson for Pituitary is not "vectorless RAG" as a replacement for Stroma. It is the retrieval workflow: inspect document structure, select likely sections, expand local/parent context, then run product-specific reasoning.

That belongs in Pituitary because section-selection prompts, status handling, historical-vs-active preference, and governance confidence are product semantics.

Current adjacent work:

• Pituitary already uses Stroma hierarchical chunking for docs through LateChunkPolicy.
• ExpandContext(IncludeParent) is already called out as the likely measurement surface in Arm B retrieval benchmark: LLM-graded RAG with ExpandContext(IncludeParent) on/off #361.
• Stroma follow-ups exist for generic spans, outlines, and record-level aggregation: Persist generic source spans for indexed sections stroma#119, Expose outline/tree reads over section lineage stroma#120, Add record-level aggregation over search hits stroma#121.
• The 2026-05-07 complexity baseline identifies internal/analysis/doc_drift.go, internal/analysis/compliance.go, internal/index/rebuild.go, and cmd/run_command.go as high-leverage complexity hotspots. PageIndex work should reduce pressure on those areas instead of adding another cross-cutting path directly into them.

Proposed Capability

Add an outline-guided retrieval path for governance analysis commands such as review-spec, check-doc-drift, and check-compliance.

A first implementation can be simple:

1. Search candidate records/chunks through the existing Stroma-backed retrieval path.
2. Read the candidate record outline/tree once Stroma exposes it, or use current section lineage where available.
3. Select relevant sections deterministically first, with optional model-assisted selection only when the analysis runtime is configured.
4. Expand selected chunks with ExpandContext, including parent context when available.
5. Pass bounded, provenance-rich contexts into the governance analyzer.

Complexity Reduction Constraint

This issue should introduce a small shared retrieval/context layer rather than wiring outline selection separately into each analysis command.

The implementation should make the PageIndex-inspired workflow explicit as a reusable phase boundary:

• candidate record/chunk search;
• outline/tree projection;
• deterministic or bounded model-assisted section selection;
• context expansion;
• provenance-rich context payload assembly.

check-doc-drift, check-compliance, and review-spec should consume that layer through a narrow API. Avoid expanding the existing analysis monoliths with more embedded search/outline/expansion branching. If the first implementation touches only one command, the new code should still be shaped so the next command can reuse it without copying retrieval logic.

Non-Goals

• No PageIndex dependency.
• No generic chat workflow.
• No MCTS/value-search implementation in this issue.
• No Stroma schema changes from Pituitary.
• No PDF/OCR extraction here.
• No MCP-specific retrieval implementation; MCP exposure belongs in Expose document structure and context expansion as MCP primitives #386 after the shared core path exists.

Acceptance Criteria

• At least one governance command can use the outline-guided context path behind a flag or config option.
• Results include enough provenance to debug selection: record ref, chunk id, heading, snapshot fingerprint, and selected/expanded context boundaries.
• The path works without an LLM when deterministic section selection is used.
• Optional LLM section selection is bounded and falls back cleanly.
• The implementation is benchmarkable by Arm B retrieval benchmark: LLM-graded RAG with ExpandContext(IncludeParent) on/off #361.
• A reusable retrieval/context abstraction exists below command and MCP transport layers, with tests at that boundary.
• The PR compares complexity against the 2026-05-07 baseline and explicitly states whether PageIndex work reduced, preserved, or increased complexity in the touched hotspots.