Fix: allow --select-model to plan a model deletion

[viniciusnunest]

Description

Fixes #5741

Right now if you delete a model file locally and try to plan its removal with
--select-model=<deleted-model>, you get:

PlanError: Selector did not return any models.

The problem is that expand_model_selections only looks at local models. Since
the file is gone, it finds nothing, backfill_models ends up empty, and
_plan() raises. The workaround was adding some unrelated active model to the
selection, which works but it is not great.

The thing is, the underlying machinery already handles deletions fine. When
backfill_models has at least one entry, deleted models show up correctly in
context_diff.removed_snapshots. So the fix is about making the selector and
plan validation aware that "no local matches" can still be valid when matching
models pending deletion in the deployed environment.

Changes:

• Extracted _load_env_models() from select_models() to avoid duplicating
  the environment loading logic
• Added expand_model_selections_with_env() that expands selections against
  both local and environment models
• In _plan(), when backfill_models is empty, it now checks whether the
  selection matched models in the environment that were deleted locally. If so,
  PlanError is suppressed and the deletions surface via removed_snapshots
• Selecting a model that exists neither locally nor in the environment still
  raises PlanError as before

Test Plan

• test_expand_model_selections_with_env: deleted models found in env, local
  models still work, wildcards match, non-existent returns empty
• test_expand_model_selections_with_env_fallback: fallback environment is used
  when target env is missing
• test_expand_model_selections_with_env_expired: expired environments ignored
• test_plan_select_model_deleted_model: end-to-end with the sushi example,
  deploys to prod, deletes model file, reloads, runs plan and checks that the
  model appears in removed_snapshots
• test_plan_select_model_deleted_model_still_rejects_nonexistent: negative
  case, makes sure garbage input still fails

Checklist

• I have run make style and fixed any issues
• I have added tests for my changes (if applicable)
• All existing tests pass (make fast-test)
• My commits are signed off (git commit -s) per the DCO

[eakmanrq]

Thanks for your submission @viniciusnunest. Does this solution assume that the selector only contains deleted models or not deleted models? What if there was a mix?

[viniciusnunest]

@eakmanrq Good question. It handles both cases.

The new code path (if not backfill_models) only kicks in when the selection matched zero local models, so when every selected model is a deletion. In that scenario, expand_model_selections_with_env confirms the models exist in the deployed environment, and we let the plan proceed so the deletions surface through context_diff.removed_snapshots.

When there is a mix (some active + some deleted), expand_model_selections still finds the active ones and backfill_models is not empty, so the new block is skipped entirely. The deleted models show up in removed_snapshots naturally because they exist in the environment but not in the local project. That part of the behavior was already working before this PR, the issue was only when the selection contained nothing but deletions.

[eakmanrq]

Is this what you need since it uses the env models? I'm wondering if you could return this alongside the current UniqueKeyDict and then use that and remove the need for expand_model_selections_with_env? Ex:

models_override, selected_fqns = model_selector.select_models(...)
if not backfill_models:
    selected_deletion_fqns = selected_fqns - set(self._models)

[eakmanrq]

@eakmanrq Good question. It handles both cases.

The new code path (if not backfill_models) only kicks in when the selection matched zero local models, so when every selected model is a deletion. In that scenario, expand_model_selections_with_env confirms the models exist in the deployed environment, and we let the plan proceed so the deletions surface through context_diff.removed_snapshots.

When there is a mix (some active + some deleted), expand_model_selections still finds the active ones and backfill_models is not empty, so the new block is skipped entirely. The deleted models show up in removed_snapshots naturally because they exist in the environment but not in the local project. That part of the behavior was already working before this PR, the issue was only when the selection contained nothing but deletions.

Thanks I understand now. Can you make sure tests cover this case?

[viniciusnunest]

Good call, refactored. select_models now returns the matched FQNs alongside the model dict, so context._plan can figure out which selections are deletions without loading env models a second time. Removed expand_model_selections_with_env entirely.

Also added test coverage for the mixed case (active + deleted in the same selection). The mixed scenario was already working before but now it is explicitly tested, test_select_models_returns_selected_fqns covers deleted-only, local-only, mixed, wildcard, and nonexistent.