refactor(learning): VRE as knowledge linter; integrators own the loop

[anormang1992]

Summary

Removes VRE's auto-learning loop orchestration. VRE now identifies gaps and persists validated fills; loop control, decision tracking, and metrics are the integrator's responsibility. The langchain demo's new `learn_gaps` tool is the reference implementation.

• Core API shrinks: `learn_all`, `on_learn`, `LearningCallback`, `CandidateDecision`, `LearningResult`, learning metrics, and the invisible `_learn_missing_depths` nested-callback are all gone
• Engine validates and persists: `LearningEngine.learn_gap(gap, candidate, source=LEARNED)` replaces callback-driven `learn_at`; field-presence and gap-relationship validation moves onto candidate models via `validate_for_gap`
• Reachability prerequisites are explicit: `reachability_prerequisites()` returns `DepthGap` objects the integrator must fill before edge placement (replaces the silent nested callback). `ReachabilityCandidate` gains an explicit `source_name` so the integrator can place edges in either direction; at least one side must match the gap primitive
• `contiguous_max_depth` is a property on `Primitive` — shared by grounding and learning engines (was duplicated)
• Demo update: new `learn_gaps` tool with integrator-owned loop, system prompt rewritten to be explicit about the block→learn→retry pattern, learner uses `think=False` for clean structured output on thinking models

Test plan

• `ruff check` clean
• 239 tests pass; coverage 74% (above 70% threshold)
• Manual end-to-end: `create test.txt` exercised depth + reachability + cascading relational gap fills via the agent and learner
• Manual end-to-end: `delete test.txt` exercised reachability prerequisites flow (depth gaps surfaced before edge placement)

🤖 Generated with Claude Code