fix(mcp-server): pass the resolved model to run-harness like scheduled runs

[aaronjmars]

What changed

The local MCP server (apps/mcp-server) never passed --model to run-harness, so a local run always used the CLI's own default model, even for skills pinned to a model in aeon.yml (for example sc-audit and deploy-uni-hook on claude-opus-5-5). Scheduled runs pass the model, so the two paths disagreed.

The server now picks the model the same way .github/workflows/aeon.yml does:

1. AEON_MODEL env (the local stand-in for the workflow's model dispatch input, next to the existing AEON_HARNESS; (config default) counts as unset)
2. the skill's own model: "..." in its aeon.yml entry
3. the global model: in aeon.yml
4. claude-sonnet-5-5

And it passes it per harness the way the workflow does:

• claude: always --model <id>.
• grok: a claude id, default or empty becomes grok-4.7, and only a grok-* id is passed.
• codex / pi / vibe / kimi / fx / cursor / hermes: --model is the MODEL_ARG from scripts/resolve-harness.sh, the same script the workflow's "Resolve harness" step runs. It maps the pick to the harness's own id, or leaves it out so the harness's default applies, based on which auth keys are set.

The per-skill aeon.yml reader now follows block-shaped entries and strips comments, a port of scripts/skill_entry.sh (which the workflow and resolve-harness.sh already use), so a block entry's model: or harness: is not missed. The run log line now names the model, and the README documents AEON_MODEL and the order.

How it was verified

• npm run typecheck and npm run build in apps/mcp-server.
• No-cost argv test: fake claude, grok and codex binaries first on PATH that save their argv and print a valid result, then tools/call over stdio against the built server with a clean env:
  • aeon-sc-audit (pinned in aeon.yml): claude got --model claude-opus-5-5.
  • aeon-heartbeat (no pin): claude got --model claude-sonnet-5-5.
  • AEON_MODEL=claude-opus-5-5 on heartbeat: --model claude-opus-5-5.
  • AEON_HARNESS=grok on sc-audit: --model grok-4.7; with AEON_MODEL=grok-4.6: --model grok-4.6.
  • AEON_HARNESS=codex, no auth keys: --model openai/gpt-6-luna; with OPENAI_API_KEY: no --model; with that key and AEON_MODEL=openai/gpt-6.1-sol: --model gpt-6.1-sol. All match what resolve-harness.sh gives the workflow.
• The per-skill model read matched the workflow's skill_entry.sh + sed pipeline for all 89 entries in aeon.yml and a fixture with block entries and trailing comments.

No real (paid) run was done.