Add KnowledgeAcquisitionModule: OOD detection, token-space fallback, and knowledge sharing

Implements the novel knowledge acquisition loop from the survey
(novel_knowledge_verification_survey.md) as a new Phase C module:

- OODDetector: two-signal gate combining JEPA predictive variance (proxy
  for VJEPA σ²) and FAISS/torch KNN distance to distinguish epistemic
  uncertainty ('I don't know this') from aleatoric uncertainty ('inherently
  ambiguous').  Adds < 1 ms overhead to the reasoning loop.

- TokenSpaceFallback: provider-agnostic bridge to a token-space LLM
  (Claude, GPT, local model).  Re-anchors the response into embedding
  space via the VLJEPA text encoder so results can be stored and shared.

- KnowledgeStore: validated embedding index backed by FAISS (with
  torch brute-force fallback).  Applies a two-layer gate at insertion:
    Layer 1 (novelty):    KNN distance check — rejects redundant contributions
    Layer 2 (consistency): contradiction detection against existing entries

- KnowledgeSharingProtocol: four-layer verification stack
    Layer 1–2: novelty + consistency (as above)
    Layer 3:   probabilistic spot-checking at 7% rate
    Layer 4:   challenge window (24 h, tracked by metadata deadline)
  TraceRank reputation scoring propagates through knowledge usage flows.
  Sybil resistance via stake-weighted validation and commit-reveal attestations.

- KnowledgeAcquisitionEngine: main coordinator routing queries across
  KNOWLEDGE_CACHE → NETWORK (Phase C stub) → TOKEN_SPACE paths.
  Tracks provenance for attribution and economic reward calculation.

Key papers:
  - Sun et al. (2022): KNN OOD detection, ICML 2022, arXiv:2204.06507
  - VJEPA (2026): variational JEPA uncertainty, arXiv:2601.14354
  - TraceRank (Shi et al., 2025): reputation via usage flows, arXiv:2510.27554
  - RIFT (2025–26): commit-reveal + peer review for AI verification

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: EightRice