Add referenced_artifacts and checkpoint tool

[seamus-brady]

Summary

PR 2 of the sub-agent-resilience plan. Implements Fixes 3 + 4: parent → child structural-context handoff via referenced_artifacts, and a lighter checkpoint tool plus skill discipline updates so agents save in chunks during synthesis instead of trying to fit it all in one final response.

Together these address the redundant-bootstrap and lost-work-on-cap patterns seen in the 2026-04-26 Nemo session, where 13 researcher delegations each re-discovered the same 309-section book and one writer delegation produced zero words after blowing its token cap.

What changed

Fix 3 — referenced_artifacts on agent_* tool calls

• New parameter on every agent_* tool: referenced_artifacts (comma-separated artifact IDs).
• Framework intercepts in dispatch_single_agent (src/agent/cognitive/agents.gleam): retrieves each artifact's content via the librarian, renders as <reference_artifact id="...">CONTENT</reference_artifact> blocks, prepends to the agent's first message.
• Resolution failures render status="not_found" markers; over-50KB bundle renders status="elided" reason="bundle_size" markers — agent always sees what was attempted, never silently dropped.
• New helpers parse_referenced_artifacts_csv and render_referenced_artifacts_bundle (both pub for testability).

Fix 4 — checkpoint tool + skill discipline

• New checkpoint(label, content) tool in src/tools/artifacts.gleam. Lighter than store_result — auto-fills tool="checkpoint", uses label as the summary, otherwise behaves identically.
• Wired into writer + researcher routing predicates and executors. Both agents can now call it.
• Updated .springdrift_example/skills/document-library/SKILL.md with explicit guidance on the checkpointing pattern (save sections as you produce them, don't assemble the whole final output in one response) and the reconnaissance-then-followups pattern using referenced_artifacts.

Test plan

• gleam build clean, no warnings
• gleam test — 2074 passed (15 new), no failures
• Pure tests on parse_referenced_artifacts_csv (3) — empty / present / invalid-json fallbacks
• Bundle rendering tests (5) — empty CSV, no librarian, valid artifact resolves with content, missing artifact gets not_found marker, multiple artifacts render in order, whitespace in CSV is trimmed
• Checkpoint dispatch tests (4) — happy path returns artifact_id, label becomes summary, missing label rejected, missing content rejected
• Routing tests (2) — writer.routes_tool("checkpoint") and researcher.routes_tool("checkpoint") both True (regression guard for the dispatch-routing bug class fixed in PR Fix three agent-tool dispatch bugs of the same family #163)
• Operator: dispatch a synthesis task with reconnaissance pattern (agent_researcher returns an artifact_id in its summary, then dispatch parallel followups with referenced_artifacts: <that_id>) and confirm the followups receive the content as a <reference_artifact> block in their first message

Notes for review

• referenced_artifacts is a string parameter (comma-separated), not a JSON array. Gleam's tool builder doesn't have an array param helper yet, and string-with-CSV parsing is robust enough for this use. The description on the tool param explicitly tells the LLM to comma-separate.
• The 50KB bundle cap is a constant in agents.gleam (referenced_artifacts_bundle_cap_chars). If operators routinely need bigger bundles, consider lifting to config; for now it matches the existing artifact-record-truncation cap of the same size.
• The existing artifact_id (singular) param remains unchanged — embeds just the ID as a hint that the agent must retrieve_result itself. referenced_artifacts is the new auto-prepend-content path. Both can coexist.
• checkpoint tool's tool field on the ArtifactRecord is the discriminator — downstream filtering (tool == "checkpoint") can find checkpoints specifically. No new schema field needed.

What's next

PR 3 — Fix 5: codify Nemo's emergent strategies as orchestration skills (orchestration-large-inputs/SKILL.md + when-to-use-writer/SKILL.md) and add Strategy Registry seeding so fresh instances boot with the floor strategies populated.

🤖 Generated with Claude Code