v0.4.6 — two-tier setup-local hardening + per-stage model flags

outrider setup-local --two-tier hardening (#47)

Fixes (make the two-tier drafter/refiner install actually run on a target repo):

• Drafter uses: ./ is rewritten to uses: remyxai/outrider@v1 at install time (previously resolved to the target repo root → no action.yml → job failure).
• The installed runner (outrider.yml) now declares and forwards the full input set the refiner dispatches (mode, publish, start-from-ref, lead-content, staged-synthesis); GitHub previously rejected these as unknown workflow_dispatch keys, so the refiner's dispatch failed and no refinement PR opened.
• Fixed _fetch_outrider_template passing a stray leading api arg (produced gh api api repos/...), which broke every two-tier template fetch.
• Two-tier PR body + install plan now document that GitHub does not run schedule: triggers on forks (drive cadence via workflow_dispatch / an external dispatcher).

New — optional per-stage model flags:

• --drafter-model, --refiner-model, --refine-model (+ --zai-key) let an install retune any stage's model — including routing it at z.ai's GLM — without hand-editing the workflow files. Provider is inferred from the model name (glm-*→z.ai, claude-*→Anthropic). Defaults are unchanged: omit the flags and all three stages stay single-provider on Anthropic (only ANTHROPIC_API_KEY needed).

Validated end-to-end on a live fork: install, uses:./→@v1 rewrite, runner accepting the refiner's dispatch inputs, and GLM routing all confirmed; 127 tests pass.