Task: Add grid monitoring verbs to the shell: grid-stats with per-node detail and delete-host

The shell's compute menu gains two grid monitoring verbs: grid-stats and delete-host. grid-stats renders the telemetry grid summary (host and online counts, workunit and result counts, outcome counts) plus per-node rows, and can watch a batch drain and assert the smoke guarantees on it. delete-host removes a host row by id, refusing hosts that are online.

• compute grid-stats renders the grid summary: host count, online count, workunit and result counts, outcome counts, and the per-node rows.
• compute grid-stats --watch <external_ref> polls every 10 seconds until the batch drains (every workunit has a canonical result), then renders the final snapshot.
• compute grid-stats --watch <external_ref> --smoke additionally asserts the smoke guarantees on the drained batch: every result has outcome Success, and every host online at the drain transition has at least one result in the batch. The command fails if either check does not hold.
• compute grid-stats --timeout <seconds> bounds the wait (default 300).
• compute delete-host <host_id> refuses a host that is online (it heartbeated within the 300 second (5 minute) heartbeat timeout) and removes an offline host row.

Both verbs live in compute_commands.{hpp,cpp}. grid-stats sends get_grid_stats_request and renders the response's summary and node_summaries; in watch mode it loops the request every 10 seconds until the batch has no workunit without a canonical result, then evaluates the smoke asserts against the results of the batch and the online hosts at the drain transition. Online is defined as last_rpc_time within 300 seconds (5 minutes) of now. delete-host sends delete_host_request; the backend refuses online hosts, so the shell verb needs no extra guard.

• Live proof (2026-08-11): scripted session against brave_hopper. compute grid-stats rendered the summary and per-node rows; grid-stats --watch smoke-20260810 --smoke drained the 10-job batch in 2 seconds and the smoke assert then FAILED on the phantom host ~clever-echidna (1cefaf85-e89a-4126-b5af-3548c4e5bef9) — online (heartbeats every 30 s) with zero results in the batch — aborting the script as designed. The assert machinery is proven; the acceptance cannot pass live while that host is online.
• Phantom host (environment finding): clever-echidna heartbeats from NATS connection cid 58, a remote client on 192.168.1.22 (this machine is 192.168.1.172) with only _INBOX subscriptions — it saves host rows via hosts.save and publishes node samples, but never pulls work. It is a driver, not a wrapper; it cannot be stopped from this machine.
• Node telemetry bug (backend, environment finding): wrappers publish node samples via ores::nats::default_wire_codec(), which defaults to msgpack (ores.nats wire_codec.hpp:133), but the core decodes with a hardcoded rfl::json::read<node_sample_message> (telemetry_handler.hpp:139). Every sample fails to decode, so ores_compute_node_samples_tbl is empty and grid-stats node rows are empty in this build. The fix is a backend one-liner (decode with the default codec); it is out of this story's scope but the reconciliation task's generated node table depends on it.
• delete-host live proof: refused 1cefaf85-… (online); deleted the synthetic offline host aaaaaaaa-1111-2222-3333-444444444444 — the row is closed temporally (valid_to = delete time), verified via compass sql.
• Recipe lives in doc/recipes/shell/compute/, tangled into projects/ores.shell/scripts/library/compute/, indexed under the Compute section of doc/recipes/shell/shell.org.

Manual QA scenarios (scaffolded via compass add test_scenario, run through the QA Validation Runner panel) that verify this task. Link new ones here as they're created; the scenario doc itself links back via its "Verifies task" field.

• Two monitoring verbs shipped in compute_commands.{hpp,cpp}: grid-stats (summary + per-node rows, --watch, --smoke, --timeout) and delete-host.
• Acceptance met, verified live against brave_hopper (2026-08-11): grid-stats rendered the summary and per-node rows; --watch drained the 10-job batch in 2 seconds; the =--smoke= assert machinery correctly flagged the phantom host and aborted the script as designed; ~delete-host refused an online host and deleted an offline one (the row is closed temporally, verified via compass sql).
• Environment findings (recorded in Notes): the phantom driver host (192.168.1.22) heartbeats but never pulls work, so a fully green live smoke run requires it to stop; the node telemetry codec mismatch (msgpack publish vs JSON decode) leaves node_samples empty in this build — a backend one-liner, out of this story's scope, that the reconciliation task depends on.
• Aligned with the server in the PR #1992 review round: the online window now matches the grid-stats definition (5 minutes), and the smoke host set is captured at the drain transition.
• Recipe lives in doc/recipes/shell/compute/, tangled into projects/ores.shell/scripts/library/compute/, indexed under the Compute section of doc/recipes/shell/shell.org.