Problem

When the main model is configured with AWS Bedrock (provider: bedrock),
self.provider == "bedrock" was being passed unconditionally to
get_model_context_length() for the auxiliary compression model.

get_model_context_length has a hard-intercept at step 1b:

if provider == "bedrock" or base_url.startswith("bedrock-runtime."):
    return get_bedrock_context_length(model)  # short-circuits all other resolution

A non-Bedrock aux model (e.g. deepseek/deepseek-v4-flash on OpenRouter) is
not in the Bedrock static table, so it falls back to
BEDROCK_DEFAULT_CONTEXT_LENGTH = 128_000. This triggers a false feasibility
warning every session start:

⚠ Compression model deepseek/deepseek-v4-flash (openrouter) context is 128,000 tokens,
  but the main model's compression threshold was 800,000 tokens.
  Auto-lowered this session's threshold to 128,000 tokens so compression can run.

…even though deepseek/deepseek-v4-flash natively supports 1,048,576 tokens.

Root Cause

_aux_cfg_provider (the aux model's own provider, resolved via
_resolve_task_provider_model("compression")) was already being fetched just
above the call site — but never passed to get_model_context_length(). Instead,
the main model's self.provider was passed unconditionally.

The invariant that provider must belong to the model being queried — not the
caller's model — is not enforced at the call site today, which is why this leak
is silent.

Fix

Pass _aux_cfg_provider when it is explicitly set (not empty or "auto"),
falling back to the main model's provider only when the aux provider is
unspecified.

# Before
provider=getattr(self, "provider", ""),

# After
provider=(_aux_cfg_provider if _aux_cfg_provider and _aux_cfg_provider != "auto" else getattr(self, "provider", "")),

Impact

The Bedrock hard-intercept is the most visible trigger, but the root cause is
general: any provider whose resolution path diverges at step 1b (Bedrock static
table) or later steps will produce wrong results when a foreign provider is
passed for an aux model.

Affected: main model = Bedrock + aux compression model on any non-Bedrock
provider (OpenRouter, OpenAI, Anthropic, etc.). Other combinations are not
affected today because non-Bedrock providers don't short-circuit at step 1b.

Functional impact is minor (threshold auto-corrects for the session), but the
warning is misleading, fires every session, and causes users to misconfigure
their setup to suppress it.

Why This Is Not a Duplicate of #17460

PR #17460 fixes get_model_context_length() to auto-load custom_providers
when the parameter is None — protecting callers that never passed the
argument. This PR fixes a different parameter on the same call: provider
is present at every call site but is being set to the wrong value.

The two fixes are complementary and non-overlapping:

Neither PR makes the other redundant. Both should land.

Testing

All 16 existing compression feasibility tests pass unchanged.

Closes #12977
Related: #13807, #17460