follow-up: automated-review workflow passes unsupported inputs to claude-code-action

[carlos-alm]

Noticed while sweeping PR #1787.

Finding: The automated-review job (.github/workflows/claude.yml) passes model and direct_prompt inputs to anthropics/claude-code-action, but the action no longer accepts them — every run emits two warning annotations:

Unexpected input(s) 'model', 'direct_prompt', valid inputs are ['trigger_phrase', 'assignee_trigger', 'label_trigger', 'base_branch', 'branch_prefix', 'branch_name_template', 'allowed_bots', 'allowed_non_write_users', 'include_comments_by_actor', 'exclude_comments_by_actor', 'prompt', 'settings', 'anthropic_api_key', 'claude_code_oauth_token', ...]

Seen at: https://github.com/optave/ops-codegraph-tool/actions/runs/28700265455/job/85116992130 (check run 85116992130, annotations).

The job still reports success so it isn't currently blocking CI, but the inputs are silently ignored, which likely means the intended prompt/model overrides aren't taking effect.

Context — why this is out of scope for #1787: PR #1787 is a pure decomposition refactor of extractors/javascript.ts, build-edges.ts, native-orchestrator.ts, and remote.ts. It does not touch .github/workflows/, and this same warning would appear on any PR in the repo (it's a config-drift issue against the current claude-code-action version, not something introduced by this change).

Suggested fix: Update .github/workflows/claude.yml to use the current supported input names (e.g. prompt instead of direct_prompt, and whatever mechanism the action now uses for model selection, if any) at lines 55-56 and 214.