consult: allow specifying model version (e.g. gemini-2.5-pro vs gemini-3-pro-preview)

[amrmelsayed]

Problem

consult -m <model> only accepts a model family (gemini, codex, claude, hermes). The specific version is hardcoded in the source, so callers cannot select a different version per invocation.

For example, gemini is pinned to gemini-3-pro-preview here:

// packages/codev/src/commands/consult/index.ts (compiled to dist/commands/consult/index.js:22)
gemini: { cli: 'gemini', args: ['--model', 'gemini-3-pro-preview'], envVar: 'GEMINI_SYSTEM_MD' },

There's no way to ask consult to run against gemini-2.5-pro (or any other version) without editing the source and rebuilding.

Use cases

• Comparing review quality across model versions (e.g. 2.5 vs 3 pro) without rebuilding
• Falling back to an older/cheaper version when the pinned preview is rate-limited or unavailable
• Pinning a specific stable version for reproducible reviews while the default tracks preview

Proposed solution

Add a passthrough flag, e.g. --model-version <id>, that overrides the hardcoded --model arg in the per-family config:

consult -m gemini --model-version gemini-2.5-pro --type spec --issue 42
consult -m codex  --model-version gpt-5         --prompt "..."

Behavior:

• When omitted, current pinned default is used (no change)
• When provided, replaces the --model arg passed to the underlying CLI
• Usage/cost extraction in usage-extractor.js already knows about multiple gemini versions (gemini-3-pro, gemini-2.5-pro), so metrics should keep working

Alternatives considered

• Calling the underlying gemini / codex CLI directly — works but skips consult's review-type scaffolding, output handling, and metrics tracking.
• Editing the pinned version in source and rebuilding — not viable per-invocation.

Acceptance criteria

• consult -m gemini --model-version gemini-2.5-pro --prompt "..." invokes the gemini CLI with --model gemini-2.5-pro
• Same flag works for codex and claude families where the underlying CLI supports it
• Omitting the flag preserves current behavior exactly
• Metrics in consult stats correctly attribute the invocation to the actual version used
• consult -m gemini --help documents the new flag