feat: audience attribute (human | agent) on technique output declarations (#224 V4)

[m2ux]

Summary

Adds an optional audience attribute (human | agent) to technique output / #### artifact declarations so each artifact's intended reader — and therefore its on-disk format — is known at authoring time. Purely additive and behaviour-neutral (absent ⇒ human), this is PR #1 of epic #224 (cluster 2 / backlog item V4) and the RC4 enabler for converting agent-state artifacts to compact structured data.

🐛 Issue 📐 Engineering

---

Motivation

Technique output declarations carry no signal about who each produced artifact is for. Human-facing artifacts (reviews, plans) and agent-only state artifacts are declared identically, so every artifact is authored as prose — verbose to produce, costly in tokens, and unreliable for an agent to read back later in the same run. Without an explicit reader marker, the downstream verbosity-reduction work (V5+) has no principled basis for deciding which artifacts may become compact structured data.

This change introduces the audience marker as the enabling primitive. On its own it changes nothing a reader sees and no workflow behaviour, but it lets later work serialize agent-audience artifacts as JSON with confidence about the intended reader.

---

Changes

• Protocol specification - documents the audience attribute on output / #### artifact declarations, the agent-⇒-JSON on-disk convention under the artifactPrefix rule, and authoring guidance (decision test, per-side anti-patterns, V5 out-of-scope note).
• Schema - adds the optional audience enum (human/agent) to the technique output schema, validated under strict parsing so an out-of-set value is rejected loudly at load.
• Loader - generalizes subsection parsing to a reserved-key set and lands audience on parsed outputs, passed through verbatim.
• Delivery projection - carries the audience through the worker-delivered technique bundle and the get_activity artifacts contract, so an artifact's format is known at write time; the projection round-trip needed no source edit.
• Corpus lint - a standalone check:audience guard flags any agent-audience output whose artifact name is not JSON, baseline-snapshotted (empty by design) and wired into the lint suite with a regression guard.
• Tests - per-surface coverage across loader parse (human/agent/absent/invalid), schema accept/reject under strict parsing, projection carry-through, and activity-contract carry-through; full suite 579 passed, 0 failed.

---

🤖 AI Assistance

• Assistant / Model: claude / claude-opus-4-8
• Context scope: repo-only
• Prompt classes: code-generation, test-writing, docs
• Provenance log: 08-provenance-log.md

---

📌 Submission Checklist

• Changes are backward-compatible (or flagged if breaking)
• Pull request description explains why the change is needed
• Self-reviewed the diff
• I have included a change file, or skipped for this reason: additive optional field, no user-facing behaviour change
• If the changes introduce a new feature, I have bumped the node minor version
• Update documentation (if relevant)
• No new todos introduced

---

🔱 Fork Strategy

• Node Runtime Update
• Node Client Update
• Other
• N/A

---

🗹 TODO before merging

• Ready for review
• Address reviewer feedback
• Squash if needed on merge (use local signed squash-merge flow)