[experimental] Add knowledge document artifact with trust lifecycle

[mk2427]

Summary

Adds a new artifact type: knowledge documents (GPD/knowledge/NNN-slug.md) with a trust lifecycle (Draft → Under Review → Stable → Superseded).

GPD's RESEARCH.md is "what the agent found." A knowledge doc is "what we've verified is correct." No existing GPD artifact has a trust lifecycle.

Why

In real physics research with AI agents, wrong understanding of the problem setup causes more wasted computation than bugs in the solver. A knowledge doc forces the agent to articulate its understanding in a reviewable form before computing.

Changes (5 files, 248 lines added)

• specs/templates/knowledge.md — template with frontmatter schema and structured sections
• commands/digest-knowledge.md — /gpd:digest-knowledge <topic or arXiv-ID>
• specs/workflows/digest-knowledge.md — light path: research → user review → Stable
• core/constants.py — KNOWLEDGE_DIR_NAME + ProjectLayout.knowledge_dir (+9 lines)
• registry.py — skill category entry (+1 line)

What it does NOT change

• No config.py changes
• No results.py changes
• No existing behavior modified
• No new agents
• No new dependencies

Future phases (separate PRs)

• Phase 2: Dependency tracking (knowledge_deps on results/plans)
• Phase 3: Gated execution (knowledge_gate: off | warn | block)
• Phase 4: Adversarial review path (opt-in, user-approved)

Test plan

• uv run ruff check clean
• 75 metadata/config tests pass
• No overlap with recent PRs (Harden staged new-project wiring #99-108)