ConTrust: credibility-weighted context priors (streamlined)

[muhaiminsk]

Streamlined branch as discussed — the prior classification and consensus
scoring only, so this can be reviewed and merged independently of the
research branch (#33).

FactReasoner enters every retrieved context into the Markov network at a
fixed PRIOR_PROB_CONTEXT = 0.9. This adds an optional per-source weight:

w = (1 − β)·prior + β·r

prior  published media-credibility rating (MBFC), exact-TLD rule for
       .gov/.edu/.int/.mil → 0.90, unrated → 0.50 neutral
r      (1 + agreed)/(2 + seen) — Beta posterior mean over agreement with
       the credibility-weighted consensus of the other evidence
β      min(a/2, 0.7) — the prior dominates until a source has a record

No existing file is modified. The scorer sets context.probability
between pipeline.build() and pipeline.score(), which the assessor
already reads. It does not touch the NLI path — it only reads
relation.probability after the fact.

Five files, all new:

• src/fact_reasoner/core/contrust.py (234 lines)
• data/priors/mbfc_idiap.csv — MBFC ratings via idiap (Apache-2.0),
  Sánchez-Cortés et al., CLEF 2024
• tests/test_contrust.py — 5 tests
• docs/examples/assessors/ex_factreasoner_contrust.py — the existing
  example with a 3-line insertion
• examples/contrust_example.py — standalone demo, no API required

Evaluations, ablations, baselines and frozen reproduction artifacts are on
feature/consensus-trust (#33).