fix: positively gate reasoning-effort branch, extend gpt-5 alias fallback

[WSxDemise]

What

Fixes the /model_settings reasoning-effort picker so it exposes the full,
real per-model reasoning_effort scale that OpenAI documents — including
xhigh/max where a model actually supports them — and makes
o1/o3/o4-mini/codex-mini-latest support reasoning_effort at all
(they had none before this change).

Why

• The picker's xhigh/max options were gated behind flags that no code
  path ever set, so no model — not even ones that genuinely support
  max — could ever show them.
• o1/o3/o4-mini/codex-mini-latest had zero reasoning_effort support
  in the UI, the catalog builder, or the API request itself, even though
  OpenAI's docs list low/medium/high for all of them.
• Once reasoning_effort was wired into the UI, a follow-up gap surfaced:
  model_factory.make_model_settings() never actually forwarded the chosen
  value to the OpenAI API for any non-GPT-5 model — the setting was visible
  and selectable but silently dropped on the wire.

How

• Added model_utils.get_openai_reasoning_effort_choices() as the single
  source of truth for which reasoning_effort values each OpenAI model
  family actually supports (per OpenAI's published docs/OpenAPI spec),
  matched via the existing boundary-aware _matches_model_tag() helper so
  short family tokens (o1, o3) can't false-positive on unrelated
  aliases (e.g. a custom model called zoo1-claude) while still matching
  legitimate embedded patterns like codex-gpt-5.
• Wired that helper into:
  • config.model_supports_setting() — so /model_settings recognizes the
    setting dynamically, even for catalog entries that never declared it
    explicitly (checking both the catalog key and the underlying
    model_config["name"] alias).
  • command_line/model_settings_menu.py — so the picker offers exactly the
    choices that model supports.
  • command_line/add_model_menu.py — so newly added models get the right
    catalog defaults.
  • model_factory.make_model_settings() — a new branch that actually
    forwards the chosen reasoning_effort to openai_reasoning_effort for
    non-GPT-5 OpenAI reasoning models, normalizing legacy aliases
    (minimal → none, ultra → max). Gated on a positive
    OpenAI-compatible-type whitelist (mirroring the existing
    _is_anthropic_model pattern) rather than a "not Anthropic" negative,
    so an aliased non-OpenAI catalog entry can't be misrouted into this
    branch either.
  • Extended the pre-existing GPT-5 branch's own condition with the same
    catalog-name alias fallback, so an aliased gpt-5* custom endpoint
    still gets full verbosity/summary/Responses-API handling instead of
    falling into the plainer generic branch.

Testing

• Full test suite: 7572 passed, 28 skipped, 1 xpassed.
• ruff check . and ruff format --check .: clean.
• New/updated targeted regression coverage in tests/test_config_full_coverage.py,
  tests/command_line/test_model_settings_menu_coverage.py, and
  tests/test_custom_model_settings.py covering:
  • Per-model choice correctness (plain gpt-5 excludes xhigh/max,
    gpt-5.6 includes max, o-series gets low/medium/high,
    fixed-effort models like o1-mini/gpt-5-pro get no choices).
  • Dynamic model_supports_setting() detection without a catalog entry.
  • Actual forwarding of reasoning_effort → openai_reasoning_effort on
    the wire for o-series/codex-mini, with legacy alias normalization.
  • Alias fallback (catalog key vs. model_config["name"]) in both the
    settings-menu and make_model_settings() code paths.
  • Boundary-safe matching: codex-gpt-5 still matches the gpt-5 bucket;
    an unrelated alias like zoo1-claude or a gemini-typed alias like
    team-o1-eval does not get hijacked into OpenAI-specific handling.

Scope

• code_puppy/model_utils.py
• code_puppy/config.py
• code_puppy/model_factory.py
• code_puppy/command_line/model_settings_menu.py
• code_puppy/command_line/add_model_menu.py
• Corresponding test files under tests/

No dependency, lockfile, or CI/tooling config changes.