fix(bench): run the network canary on the host PATH

[e54-bot]

Summary

Under network off, --canary-cmd ran with the agent's PATH, where every network tool resolves to a blocker shim. A canary calling a blocked tool (for example curl -sf <url>) exited non-zero whether or not the network was reachable, so the run recorded fetch-blocked+canary-checked without any real network check.

The canary now runs with BENCH_HOST_PATH (the unscrubbed host PATH the harness already exposes), so it fails only when the network is actually unreachable, and a reachable network marks the run invalid as documented. The marker, --canary-cmd help, and both benchmark docs now describe this accurately.

Fixes #590

Test plan

• test_network_canary_does_not_resolve_a_blocker: a canary that inspects command -v curl no longer resolves the blocker dir (fails on the old implementation, passes now)
• test_network_canary_calls_a_real_tool_on_the_host_path: a planted curl on the host PATH that succeeds marks the run invalid (network reachable); one that fails records fetch-blocked+canary-checked
• python3 -m unittest test_agent_bench — new tests pass; the only failures (test_scenarios_are_solvable_and_not_vacuous, test_matrix_skips_...) reproduce identically on unmodified main with the local debug binary (suite calibration vs binary version), unrelated to this change

[e54-bot]

LGTM — verified against issue #590's acceptance criteria.

• The canary now runs with PATH=BENCH_HOST_PATH while keeping the rest of the scrubbed env, so it resolves real network tools instead of the blocker shims; BENCH_HOST_PATH is always populated by build_env (agent_bench.py:259) and the only caller passes that env (agent_bench.py:488). Marker recording at agent_bench.py:510 and the docs (agent-benchmark.md, agent-benchmark-howto.md) now match the behavior.
• Both new tests fail under the previous implementation and are deterministic across hosts (the command -v case falls through to exit 0 whether or not the host has curl; the end-to-end test uses a local fake curl, no real network).
• The two local suite failures (test_scenarios_are_solvable_and_not_vacuous, test_matrix_...) reproduce identically on unmodified main and are unrelated.

Note: submitted as a comment because e54-bot authored this PR and GitHub does not allow self-approval; the verdict itself is an approve.