v0.16.0 — Task-Typed Fusion & Fan-Out Safety

The Tier-2/3 half of the same 8-category arXiv sweep that produced v0.15.0 (M18) — six integrations refining the aggregation and safety layers rather than the verifier. All opt-in / degrade-safe, model-free where possible, each mapping onto an existing component.

Added

• Task-typed fusion aggregation (CHIMERA_FUSION_TASK_TYPED) — vote on a logic/single-answer task when the panel reaches a clear majority (a correct minority isn't synthesized away), synthesize on knowledge tasks. [MALLM 2607.05477]
• Panel-independence metric + blind-panel guard — the fusion panel was already blind by construction; this formalizes it with a regression guard and FusionTrace.panel_diversity(). [2607.02507]
• TraceProbe anti-pattern detectors — cheap search-loop / verification-skip signals wired into failure retry-feedback (auditable "why it failed"), never a hard gate. [2607.06184]
• Bug-report task normalizer (solve --normalize-task) — reshapes a rambling bug report into salient-facts-first form before planning; only the prompt, not the memory identity. [2607.07593]
• PROBE live wiring (solve --probe-log) — records (arm, proxy, reward) per attempt so probe-select --from-log decides escalation-pays-or-not from measured evidence. [2607.06879]
• Aggregate cross-agent monitor + collusion red-team — closes the split-exfiltration gap a per-worker monitor can't see under fan-out; escalates to review only. [FakeLab 2607.07368]

Parked (honest — not in this release)

Dyserve blast-radius weighting + residual re-solve, exponential-weights judge weighting, skeptical-executor delegation clause, hard token-cap + evict-completed-work, cache-locality routing, hot-worktree pool for the crew, and p95/p99 latency on the cost bench. None blocking.

Still alpha. Everything above defaults to prior behaviour.

Full changelog: https://github.com/brcampidelli/chimera-agent/blob/main/CHANGELOG.md