…roviders (NousResearch#12140)

``normalize_model_for_provider`` stripped a ``vendor/`` prefix from a
model name only when the prefix matched the canonical target provider
id.  Regional-variant providers (``minimax-cn``, ``kimi-coding-cn``) —
which route to the China region of the same vendor — have a different
canonical id from the base vendor, so a natural config like

    model:
      provider: minimax-cn
      default: minimax/minimax-m2.7

reached the MiniMax-CN API with the ``minimax/`` prefix intact and
produced:

    unknown model 'minimax/minimax-m2.7' (2013)

The same shape breaks ``kimi-coding-cn`` for ``kimi/…`` /
``moonshot/…`` prefixed IDs.  Reporter: NousResearch#12140.

Root cause
----------
``_strip_matching_provider_prefix`` (hermes_cli/model_normalize.py):

    normalized_prefix = _normalize_provider_alias(prefix)     # "minimax"
    normalized_target = _normalize_provider_alias(target)     # "minimax-cn"
    if normalized_prefix == normalized_target:                 # False
        return remainder
    return model_name                                          # slash preserved

The existing ``_PROVIDER_ALIASES`` map (``minimax-china``, ``minimax_cn``
→ ``minimax-cn``) handles *regional-target aliases*, but there's no
relationship expressing "the base vendor ``minimax`` is an acceptable
prefix for the regional target ``minimax-cn``".

Fix
---
Add a ``_REGIONAL_VARIANT_ROOTS`` map:

    _REGIONAL_VARIANT_ROOTS = {
        "minimax-cn": "minimax",
        "kimi-coding-cn": "kimi-coding",
    }

``_strip_matching_provider_prefix`` now tries the existing equality
check first (unchanged), then falls through to: *if target is a
regional variant and the prefix resolves to its base vendor, strip
it.*  Any other prefix-target pair is still preserved — no widening
of the strip behaviour for unrelated vendors, no symmetric
"strip ``minimax-cn/`` for ``minimax`` target" rule (verified with
an explicit test that pins that asymmetry).

Behaviour matrix (resolved after this fix)
------------------------------------------
| Input | Target | Before | After |
| --- | --- | --- | --- |
| ``minimax/minimax-m2.7`` | ``minimax-cn`` | slash preserved | ``minimax-m2.7`` |
| ``MiniMax/MiniMax-M2.7`` | ``minimax-cn`` | slash preserved | ``MiniMax-M2.7`` |
| ``kimi/kimi-k2`` | ``kimi-coding-cn`` | slash preserved | ``kimi-k2`` |
| ``moonshot/kimi-k2`` | ``kimi-coding-cn`` | slash preserved | ``kimi-k2`` |
| ``kimi-coding/kimi-k2`` | ``kimi-coding-cn`` | slash preserved | ``kimi-k2`` |
| ``arbitrary/path`` | ``minimax-cn`` | unchanged | **unchanged** |
| ``minimax/foo`` | ``minimax`` | already stripped | **unchanged** |
| ``zai/glm-5.1`` | ``zai`` | already stripped | **unchanged** |
| ``anthropic/claude-sonnet-4.6`` | ``openrouter`` | preserved (aggregator) | **unchanged** |
| ``minimax-cn/x`` | ``minimax`` | preserved (no symmetric rule) | **unchanged** |

Narrow scope — explicitly not changed
-------------------------------------
* Case normalisation.  The reporter also suggested that
  ``minimax-m2.7`` (lowercase) should be rewritten to ``MiniMax-M2.7``
  (PascalCase) for the MiniMax-CN API.  That's a config-authoring issue
  (``hermes_cli/models.py:192-200`` lists the canonical PascalCase
  form), and adding a case-lookup table for every MiniMax generation
  is scope creep.  The prefix-strip fix resolves the symptomatic
  ``unknown model 'minimax/...'`` error directly; any remaining
  case-sensitivity is a separate, smaller fix.
* ``_resolve_auto`` in ``agent/auxiliary_client.py``.  The reporter's
  "Fix 1" proposal was to normalise the model there, but the bug is
  actually at the normaliser layer — ``_resolve_auto`` *does* route
  through ``resolve_provider_client`` which already calls
  ``_normalize_resolved_model`` at every provider branch.  Fixing
  ``_strip_matching_provider_prefix`` also repairs every other caller
  of ``normalize_model_for_provider`` for regional-variant providers
  (``switch_model``, fallback-model resolution, CLI picker), not just
  auxiliary.
* launchd ``.env`` loading (reporter's "Fix 2").  Install-script /
  docs territory — mission-de-prioritised and orthogonal to this bug.

Regression coverage
-------------------
``tests/hermes_cli/test_model_normalize.py`` gets a new
``TestIssue12140RegionalVariantPrefixStripping`` class with 17
parametrised cases (3 minimax-cn + 3 kimi-coding-cn + 10 preserved-
behaviour canaries + 1 asymmetry pin).  7 of them fail on clean
``origin/main`` with the exact reporter symptom
(``assert 'minimax/minimax-m2.7' == 'minimax-m2.7'``), confirming the
regression is caught.  The remaining 10 pin preserved behaviour so a
future refactor can't accidentally widen or narrow the strip rule.

Validation
----------
``source venv/bin/activate && python -m pytest
tests/hermes_cli/test_model_normalize.py -q`` -> 72 passed (55
pre-existing + 17 new).

CI-aligned broad suite: 12909 passed, 39 skipped, 24 pre-existing
baseline failures — none in ``hermes_cli/model_normalize.py`` or its
test file.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>