feat(0.x): researcherProfile + multiHarnessResearcherFanout (1.3.0)

[tangletools]

Summary

Phase 1.A of the driven-loop substrate push. Ships
researcherProfile() + multiHarnessResearcherFanout() as a new sub-
export @tangle-network/agent-knowledge/profiles. Pairs with
@tangle-network/agent-runtime/loops — the profile owns prompt + output
adapter + validator; the runtime kernel owns iteration, concurrency,
cost, abort, and trace emission.

• Three validator invariants: namespace isolation (cross-tenant writes
  hard-fail), provenance (every item carries >= 1 evidence entry),
  citation density floor (default 0.7).
• Propose-don't-apply: the profile emits KnowledgeUpdate[]; the
  caller decides whether to route them through
  applyKnowledgeWriteBlocks / a KbStore put. The profile module
  imports no FS primitive — enforced by a test that scans the source.
• Loose output: items, citations, proposedWrites are required;
  gaps, notes, and any agent extras land in raw rather than
  getting dropped. Validator scoring:
  0.4·citation_density + 0.2·source_diversity + 0.2·recency_match + 0.2·gap_coverage.
• Three-harness default fanout: opencode/zai-coding-plan/glm-5.1,
  claude-code, codex.
• agent-builder god-object (1665 LOC) NOT touched in this PR; audit
  doc captures the extraction map for a follow-up agent-builder PR.

Substrate audit findings

Full audit: /tmp/agent-knowledge-audit.md. Summary of the three
primitives the spec demanded:

Primitive	State	This PR
Multi-tenant namespace	PARTIAL — KnowledgeFreshnessStore is per-workspaceId; KbStore + types are single-tenant	Validator enforces namespace at the profile layer; defence-in-depth. Substrate-wide promotion = follow-up PR.
Provenance + confidence + supersede + retract	PARTIAL — KnowledgeClaim has refs/confidence/status; no top-level KnowledgeItem with supersedes[] / retractedAt / authoredBy	Profile ships local KnowledgeItem whose field names match the eventual substrate promotion 1:1.
Propose + materialize	GOOD — parseKnowledgeWriteBlocks / applyKnowledgeWriteBlocks already separate proposal from materialize	Profile honours the contract; never calls materialize.

One PR (not two) because the two PARTIAL primitives don't block the
researcher work — the validator + profile-local types deliver the
spec's invariants today, and the follow-up substrate refactor will
collapse them one-for-one without changing the profile surface.

Test plan

• pnpm typecheck clean (strict + noUncheckedIndexedAccess)

• pnpm build clean (tsup ESM + DTS, new dist/profiles/ entry)

• pnpm test — 88 passed / 5 skipped (5 skipped are pre-existing
  live-HTTP tests; 24 new tests across validator + adapter +
  multi-harness + runLoop integration)

• biome check src tests — no new warnings (2 pre-existing in
  unchanged src/lint.ts + src/wikilinks.ts)

• manual smoke through dist:

  ```
  decision: pick-winner
  iterations: 2
  winner.agentRunName: researcher-stub-a
  winner.verdict.valid: true
  winner.verdict.score: 1.000
  winner proposedWrites: 1
  ```
  

Follow-ups (deferred — out of scope here)

• feat/knowledge-primitives — promote KnowledgeClaim → top-level
  KnowledgeItem substrate-wide; add namespace to KbStore get/put;
  collapse profile-local types.
• agent-builder refactor: consume researcherProfile from
  auto-research-runner.ts. Target: drop ~1000 LOC from the 1665-line
  god-object. Domain-generic loop pieces move out; persona pool +
  lineage + budget + cycle deadline + decideRepairAction stay.