feat(models): default gpt-6-astra second opinions to low reasoning effort

[datj9]

Summary

crossagent's default GPT-6 Astra second opinions now run at low reasoning effort to save cost. A second opinion is a reviewer, not an author — low scores strongly for review at roughly a third of the tokens (AA Coding Agent Index 67, above GPT-5.6 Sol). The level rides a visible -c model_reasoning_effort=low override; a user's own ~/.codex/config.toml (which may be set to high for interactive work) is never modified.

What changed

• Advisor.default_reasoning_effort + reasoning_effort_config_key — codex sets low / model_reasoning_effort; every other advisor is None (no reasoning knob → a requested level is dropped with a stderr warning, never guessed).
• effective_reasoning() — the default applies only on a fresh call whose model is the advisor's own default_model (alias-resolved, case-insensitive), so a substituted --model or a resume never forces low. An explicit --reasoning wins on both fresh and resume (mirrors --model); --reasoning default suppresses the override; an invalid level exits 2.
• --reasoning flag on both the ask and job-start paths. The effective level is persisted to the session registry and command.json.
• Escalation and verification (both fresh GPT-6 delegations) apply the same low default via a shared default_reasoning_for_model() gate — so they no longer silently run at the user's config.toml effort.
• Argv ordering: -c is emitted before the resume subcommand (codex rejects it after).

Behaviour change

Default codex asks with the default gpt-6-astra model now send -c model_reasoning_effort=low. Overridable via --reasoning <level>, --reasoning default, or advisors.json ("codex": {"default_reasoning_effort": null}). All other advisors unchanged.

Testing

• python -m pytest -q → 624 passed (was 583; +41 new tests).
• Ruff: no new findings vs main.
• Covered: reasoning_args rendering, the default-model gate (incl. alias + case + config default_model:"gpt6"), resume symmetry, --reasoning default/invalid/no-knob-warning, registry + command.json persistence, and escalation/verification argv.

Plan (with pre-implementation review notes addressed) is in docs/plans/gpt6-low-reasoning-default.md.