v0.19.4 — crew resilience

A thirteenth adversarial review, of the multi-agent roles/crew surface. Most of it reviewed clean (sequential hand-off, isolated-crew aggregation determinism, project-orchestrator termination, verifier-tied crew success, hierarchical-orchestrator crash-safety). Three fixes keep a crew honest and crash-proof.

Fixed

• A crew's parallel review no longer crashes when one reviewer fails. parallel_review / chimera crew --mode supervisor used Executor.map, which re-raises the first worker's exception — so one worker hitting a transient provider error sank the whole run and discarded every other reviewer's completed work. Each reviewer now degrades to an [error] message; the panel survives N-1 workers.
• The supervisor now sees the real consensus. Reviews were consolidated (near-duplicates merged) before the supervisor synthesized them, collapsing a 3-to-1 majority into a 1-to-1 tie; it now receives the raw reviews (dedup is kept only for memory storage).
• Per-role tool allowlist is now enforced, not decorative — Role.allowed_tools filters the registry fail-closed before the worker's agent loop, so a role can't reach a tool outside its remit.

Gate: 1455 passed / 3 skipped, mypy + ruff green.

Upgrade: `pip install --upgrade 'chimera-agent[full]'`