Imported engine.model is silently dropped when max-ai-credits is set

[crmitchelmore]

What happened

When a workflow sets max-ai-credits and takes its engine.model from an
imported file, the compiler silently discards the pin. The generated lock falls
back to repository variables:

COPILOT_MODEL: ${{ vars.GH_AW_MODEL_AGENT_COPILOT || vars.GH_AW_DEFAULT_MODEL_COPILOT || 'claude-sonnet-4.6' }}

Compilation reports 0 error(s), 0 warning(s), so nothing signals that the
workflow is no longer pinned. In a consumer repo that has not set those
variables, the run quietly executes on claude-sonnet-4.6 instead of the model
the author selected.

Either combination on its own is fine — the pin only disappears when both are
present.

Reproduction (v0.82.13)

.github/workflows/shared/model.md:

---
engine:
  id: copilot
  model: gpt-5.6-sol
---

.github/workflows/probe.md:

---
name: Probe
on:
  workflow_dispatch:
permissions: read-all
max-ai-credits: 1500
imports:
  - shared/model.md
safe-outputs:
  add-comment:
    max: 1
---

# Probe

Say hello.

$ gh aw compile probe
✓ probe.md
✓ Compiled 1 workflow(s): 0 error(s), 0 warning(s)

$ grep 'GH_AW_INFO_MODEL:' probe.lock.yml
  GH_AW_INFO_MODEL: ${{ vars.GH_AW_MODEL_AGENT_COPILOT || vars.GH_AW_DEFAULT_MODEL_COPILOT || 'claude-sonnet-4.6' }}

max-ai-credits	engine.model from	Resolved model in lock
absent	import	gpt-5.6-sol ✅
absent	inline	gpt-5.6-sol ✅
1500	inline	gpt-5.6-sol ✅
1500	import	repository-variable fallback ❌

Expected

An imported engine.model should resolve identically to an inline one. If some
combination genuinely cannot be supported, the compiler should fail rather than
emit a workflow whose model pin has silently disappeared.

Why this matters

Sharing engine: through an import is the only way to define a model default in
one place: gh-aw permits exactly one engine: across a workflow and all of its
imports, so an importing workflow cannot restate the pin as a safety net. A
catalogue of workflows that centralises its model pin this way loses the pin on
exactly the workflows that also cap their spend — and finds out at runtime.

Environment

• gh-aw v0.82.13
• engine copilot, CLI 1.0.70