hackathon/john-kioko-weighted-trust: WeightedTrust plugin (v2 — discriminator + determinism + CI)

Description

The pitch.

## WeightedTrust — recency-decayed, volume-weighted reputation (v2)

Addresses all feedback from the v1 review on #143.

---

### Reviewer feedback → fixes

**1. Non-vacuous discriminator** ✅
- Added `validate_stale_report_attacker_cheated` + `validate_stale_report_temporal_split` validators, registered in `VALIDATORS["stale_report"]`
- Added `stale_report` scenario with `HonestTrader`, `ReformedAttacker`, `TrustObserver` agent brains — the attacker behaves well for 6 rounds, then cheats for 4
- `test_stale_report_scenario.py` proves **WeightedTrust produces a lower score than ScoreAverageTrust** for the same stale-positive-then-negative evidence:
  - `score_average`: score = **0.6** (6 positives + 4 negatives → 6/10 — deceptively trusted)
  - `WeightedTrust`: score **< 0.5** (old positives decayed, recent negatives dominate)
- The core claim — that weighting beats naive averaging on stale evidence — is now **demonstrated**, not just asserted

**2. Wall-clock removed — fully deterministic** ✅
- Replaced `time.time()` with caller-supplied `set_tick(t)` + `evidence.timestamp`
- `score()` is now a **pure function** of evidence list + current tick — never reads the system clock
- Verified determinism:
  - `test_identical_scores_across_instances` — two independent instances, same inputs → identical scores
  - `test_seed_independence` — WeightedTrust uses no RNG; scores are identical across "seeds" 42, 7, 1337

**3. ci-local cleared** ✅
- `ruff check`: ALL PASS (I001 import ordering fixed)
- `ruff format`: applied to all changed files
- `pyright`: no `reportPrivateUsage` — tests use **public API only** (`score`, `report`, `attest`, `stake`, `set_tick`)

**4. Novelty vs #139** ✅
- PR #139 is `SybilResistantTrust`: 2-pass eigen-weighting for sybil/collusion defense
- WeightedTrust's **distinct** capability is **recency decay** — stale evidence loses weight exponentially. Neither `score_average`, `agent_receipts`, nor `sybil_resistant` provide this
- Dropped reporter-credibility tracking to eliminate overlap

---

### Files

**New:**
- `nest_plugins_reference/trust/weighted.py` — the plugin (deterministic, tick-based)
- `tests/test_weighted_trust.py` — 11 conformance + behaviour + determinism tests
- `tests/test_stale_report_scenario.py` — 4 discriminator tests proving weighted < average
- `scenarios/stale_report.yaml` — scenario YAML
- `nest_core/scenarios_builtin/stale_report.py` — agent brains (HonestTrader, ReformedAttacker, TrustObserver)

**Modified:**
- `nest_core/plugins.py` — registered `("trust", "weighted")`
- `nest_core/scenarios.py` — registered `"stale_report"` scenario type
- `nest_core/validators.py` — added stale_report validators

### Test results
```
52 passed in 0.22s (15 new + 37 existing, zero regressions)
ruff check: All checks passed
ruff format: clean
```

Rebased on latest main (eb58b7b).

Author: **John Kioko** (@Mutinda-Kioko)