fix(provider): wizard honors default=None — never force-writes choice/boolean values (in_memory defect)

[Brian Krabach (bkrabach)]

Root cause

_prompt_for_field's choice branch (in amplifier_app_cli/provider_config_utils.py) had no representation for "unset": it computed effective_value = existing_value or default and then hard-coded default_choice = "1". That meant a provider ConfigField declaring default=None (deliberately, to signal "leave unset — use provider/model default") with choices=[...] silently collapsed to choices[0] and always wrote a value on Enter.

Causal chain of the real-world defect

1. provider-openai declares prompt_cache_retention with default=None + choices=["in_memory", "24h"] — the field's own source comment says None means "leave unset."
2. The wizard's choice-field renderer force-selected choices[0] regardless.
3. settings.yaml got prompt_cache_retention: in_memory written on every fresh configure, even though the user never chose it.
4. gpt-5.6 models reject in_memory at runtime, producing a drop-warning every session.

The same bug force-writes text_verbosity=low for provider-openai, and would force-write any other provider's default=None choice field the moment it's declared (e.g. an anthropic reasoning_effort field shaped the same way).

A related latent bug lived in the boolean branch: default and default.lower() with default=None evaluates to None (short-circuit and), and Confirm.ask(default=None) still accepts that as a default — so pressing Enter wrote the literal string "none" into config.

Fixes (all in app-cli; no provider or core changes)

1. Choice branch (provider_config_utils.py:610-649): when default is None and the field is not required, prepend a "(leave unset - use provider/model default)" option and make it the default selection. Selecting it returns None, so the caller omits the key — matching the existing text-field "may be absent" contract. A real existing_value still takes priority as the pre-selected option (offset to account for the prepended entry). required=True fields keep the original choices[0]-default behavior unchanged.

2. Boolean branch (provider_config_utils.py:583-608): guard default is None explicitly instead of relying on and-chaining producing an accidental None. Confirm.ask is called with an explicit default=None; pressing Enter now returns None → key omitted. Never writes "none".

3. Removed dead cli_overrides machinery in configure_provider(): the model=, endpoint=, deployment=, use_azure_cli= parameters and their handling (previously ~:716-719, :767-771, :846-851, :884, :891) had zero callers anywhere in the app or tests (grep-verified across amplifier_app_cli/ and tests/). use_default_credential/use_managed_identity targeted a field no provider even declares. non_interactive and every other parameter are untouched.

4. Extracted the two byte-identical 42-line non-interactive-mode blocks (Phase 1 pre-model fields, Phase 3 post-model fields) into one _apply_non_interactive_field() helper — mechanical, behavior-preserving refactor.

Tests (tests/test_provider_config_unset_fields.py, 22 new)

• Regression: choice field default=None + required=False omits the key on Enter (the in_memory defect) instead of force-writing choices[0].
• Choice field with a real default: unchanged behavior (no unset option).
• Choice field with existing_value: preserved as the default selection, positioned after the prepended unset option.
• Boolean default=None: key omitted, never writes "none".
• Full _should_show_field predicate matrix: literal / contains: / not_contains: / startswith: / not_startswith:, missing key, multiple conditions.
• requires_model Phase-1/Phase-3 split: a requires_model field's show_when is proven to see default_model only after model selection (Phase 2) — spied via _should_show_field.
• End-to-end exact-key-set assertion for a full configure_provider() run against a mocked provider schema (asserts the omitted fields are truly absent, not just falsy).
• Dead-parameter removal: model=/endpoint=/deployment=/use_azure_cli= are no longer accepted parameters.

Test results

Full suite: 1491 passed, 1 skipped, 13 deselected, 1 xfailed (baseline 1469 passed + 22 new tests; zero regressions).

$ python3 -m pytest -q
1491 passed, 1 skipped, 13 deselected, 1 xfailed, 3 warnings in 7.02s

🤖 Generated with Amplifier

[Brian Krabach (bkrabach)]

Maintainer admin-merge: self-authored, self-approval impossible

Basis: Root cause of the reported gpt-5.6-luna prompt_cache_retention: in_memory defect. The wizard's choice renderer had no "leave unset" concept and collapsed None-defaults to choices[0]; it now renders (leave unset) as the default and omits the key entirely when left unset. Added a matching boolean None guard, and removed the dead cli_overrides machinery.

Testing: 22 new tests added covering previously-zero-coverage code paths (choice branch, show_when matrix, requires_model phase split, exact-key-set assertions). Full suite: 1491 passed.

DTU validation: Control run reproduced the defect exactly (force-wrote in_memory + text_verbosity=low; runtime emitted a "Dropping unsupported config" warning). Treatment run omitted both keys with zero warnings. Leave-unset behavior was separately verified on Anthropic's reasoning_effort field.

CI: All 9 matrix jobs + CLA green (see checks tab).

This PR is self-authored. Self-approval is not possible under branch protection, so this merge is being performed via gh pr merge --admin at explicit user direction, per the established documented maintainer admin-merge pattern (~30 PRs merged this way recently). Recorded here for auditability.