feat(verification): add --prompt-file and --model flags to the live-capture harness

Extends scripts/verification/run-live-verification.ts with two flags needed
to run the Phase 1 matrix:

  --prompt-file <path>  Load scenario.prompt from a file at runtime.
                        Required for scenarios whose prompts exceed the JSON
                        file size (e.g. long-context with ~180K tokens).
  --model <id>          Pin the model for both OG and plugin commands.
                        Substitutes {model} in templates; for templates
                        without the placeholder, auto-appends `--model <id>`.

Internal changes:
  - ParsedArgs gains optional promptFile + model
  - renderCommand(template, string) becomes renderCommand(template, options)
    with options = {prompt, model?, autoAppendModel?}
  - New resolveScenarioPrompt(scenario, {promptFile}) returns scenario.prompt
    or reads the file (throws on missing path)
  - runScenario threads args.model + args.promptFile through both commands

10 new tests (parseArgs, resolveScenarioPrompt, renderCommand template +
auto-append semantics). Full regression 1316/1316. Lint clean. Typecheck
clean. No existing behavior changes \u2014 defaults preserved.

Unblocks Phase 1 Task 1.2 (Wave 1 captures with pinned Haiku 4.5) per
.sisyphus/plans/griffin-parity-and-cc-optimizations.md Task 1.5.