feat(bench): update the llms-benchmark

[using-system]

Umbrella issue for a round of updates to the llms-benchmark: its results tables in .llms-benchmark/README.md and the /launch-llms-benchmark command that produces them.

Planned so far:

• an Effort column right after Model in both results tables; every existing row was measured at medium;
• /launch-llms-benchmark takes the effort as an optional third argument (medium by default), passed to each CLI's effort flag; model, effort and CLI together identify a row.

More changes to the benchmark will be added here as the work proceeds; the PR that closes this issue lists them all.

[using-system]

Amended choices recorded while running the first batch of this issue (gpt-6-luna and gpt-6-sol on copilot, at medium and high):

• Copilot launch line: COPILOT_TASK_WAIT_TIMEOUT_SECONDS=7200. In prompt mode Copilot waits at most 600 s for background tasks, then cancels them (session.warning background_task_wait_timeout, subagent.completed with cancelled: true). The first gpt-6-luna run dispatched observe-run in background mode and lost its whole observation ten minutes in, with no report; that run is void, not a row. /launch-llms-benchmark now sets the variable on the copilot launch line and says why. The earlier copilot rows all finished under ten minutes and never hit it, so none is invalidated.
• The oddyssey column of the new rows reads 1.13.0, on the maintainer's decision, for coherence with the table. The runs executed on 1.13.1, whose .apm/ differs from 1.13.0's by one local-stack reference file (otel-lgtm-env.md).
• Row identity: model, effort and CLI together - the same model at two efforts is two rows.

[using-system]

Further amended choices, recorded before the PR update:

• opencode's --variant is checked against the model's variants. opencode run accepts any variant name, an unknown one included (verified with --variant bogus), records it on every message and sends no reasoning effort. opencode models openrouter --verbose lists each model's variants: the z-ai/glm-5.3* models and deepseek/deepseek-v4.1-flash offer low, high and max only. Their rows, measured with --variant medium, ran at the provider's default effort, so their Effort column reads default; z-ai/glm-5.3-prime's medium slot was run without --variant and is labelled default too. The command's opencode preflight now says so.
• Copilot's --effort high is applied: subagent.configured carries reasoningEffort: high on the high runs (medium on the medium ones) and reasoning tokens rise with it.
• New rows: anthropic/claude-opus-5.5 on claude, and openai/gpt-6-luna-pro, openai/gpt-6-sol-pro, x-ai/grok-4.7, z-ai/glm-5.3-prime on opencode, each at medium (or default) and high, oddyssey column 1.13.0 as decided.

[using-system]

Row identity amended on the maintainer's decision: model, effort, CLI and provider together identify a results row (was model, effort and CLI). The command's arguments and table contract and the README say so.