Add source-grounded retrieval inspector

[WalksWithASwagger]

Context

RAG is already useful, but the user needs to see why a chunk was retrieved and whether it helped. The latest source-grounded tools make citations and source visibility table stakes.

Source: docs/WHISPERFORGE-MASTER-PLAN-2026-05-18.md (wf-retrieval-inspector, Phase 2).

Linear: BC-223.

Acceptance Criteria

• Expose retrieved KB chunks per stage with source path, heading/offset when available, score, and reason/query context.
• Show always-included voice anchors separately from semantic retrieval hits.
• Persist retrieval metadata into run artifacts so it can be inspected after the session.
• Add a compact UI or exported report that compares legacy KB dump vs RAG for the current capture/profile.
• Keep Auto/Always/Never RAG behavior compatible with current settings.

Tests/Evals

• Unit tests for retrieval metadata shape and serialization.
• Fixture test that proves voice anchors and semantic hits are distinguishable.
• No network-dependent eval required.

Verification

• make test
• make eval-fixture
• git diff --check

Agent Instructions

Build on whisperforge_core/rag instead of adding a parallel retrieval path. Keep UI density modest; this is an inspector, not a second app.

Out of Scope

Do not replace the embedding model or rewrite RAG scoring unless required by tests.

[WalksWithASwagger]

PR #26 is ready for review: https://github.com/WalksWithASwagger/spektorAI/pull/26\n\nThis slice adds retrieval hit metadata, role separation for voice anchors vs context hits, run-artifact retrieval inspector output, and UI inspection from the benchmark dialog. Verified with make test, make eval-fixture, ui_smoke, make smoke, local Playwright UI checks, git diff --check, and ops/roadmap/features.json validation.