fix: remove redundant provider label concatenation in _deduplicate_model_ids

[lost9999]

Problem

When multiple providers expose models with the same name (e.g., DeepSeek V4 Flash is available from Xiaomi, Ollama, HuggingFace, etc.), the _deduplicate_model_ids() function appends the provider name to the model label for every duplicate found. This results in garbled labels like:

Deepseek V4 Flash (Xiaomi) (Ollama) (HuggingFace) (Google-Gemini-Cli)

The label should reflect the model's own metadata, not accumulate provider names.

Root Cause

In api/config.py, the _deduplicate_model_ids function (around line 945) appends (provider_name) to model["label"] when deduplicating models with the same name. Since the function iterates across all provider groups and appends the label for every match, models exposed by many providers get increasingly garbled labels.

Fix

Remove the label concatenation logic entirely. The ID deduplication (prefixing with @provider_id:) is preserved — it's the correct way to differentiate same-named models from different providers. The label should be left as-is from the model's own metadata.

Before

model["id"] = f"@{pid}:{original_id}"
provider_name = group.get("provider", pid)
if model.get("label") != original_id:
    model["label"] = f"{model['label']} ({provider_name})"
else:
    model["label"] = f"{original_id} ({provider_name})"

After

model["id"] = f"@{pid}:{original_id}"

Testing

• Verified /api/models returns clean labels with no accumulated provider suffixes
• Model list in WebUI displays correctly with 11 clean provider groups
• No regression in model switching or provider routing

[nesquena-hermes]

Thank you @lost9999 for filing this — your bug report was the seed for v0.50.277, shipped via release #1515. Closing this PR but only because the same fix shipped at a different layer, and Co-authored-by: lost9999 on the merge commit preserves your attribution for catching it.

Why we landed the fix differently

I traced the bug from first principles before merging. The dedup function in _deduplicate_model_ids() only ever appends ONE provider name per location it visits — so it cannot accumulate four parentheticals ((Xiaomi) (Ollama) (HuggingFace) (Google-Gemini-Cli)) on its own. That made me suspicious that removing the label-suffix logic would only paper over the symptom.

I wrote a Python repro and tested three theories:

Scenario	Symptom
Independent dicts per group	One parenthetical per group ✓ (correct)
Shared dict reference across groups	N-1 parentheticals accumulate ✗ (matches vishnu's report)
Function called twice	Idempotent (the @ guard at line 930 catches it)

The shared-reference scenario reproduces the exact symptom. I then traced the 5 group-build paths in get_models_grouped() and found:

• OpenRouter / ollama-cloud / _PROVIDER_MODELS / named-custom paths all already build independent dicts ✓
• The unconfigured-provider fall-through at api/config.py:2078 shares auto_detected_models across all groups that hit that branch ✗

So when (for example) Ollama, HuggingFace, and a custom OpenAI-compat endpoint all auto-detect together, all three groups point to the SAME list with the SAME dicts inside. When dedup mutates those dicts to add @provider_id: prefixes, every group sees every other group's mutation.

The hidden bug your PR would not have fixed

Removing the label-suffix logic in _deduplicate_model_ids() makes the visible labels look clean (Deepseek V4 Flash). But the id field is corrupted by the same mechanism: all groups end up with @xiaomi:deepseek-v4-flash (whichever provider_id won the alphabetical-first race). The user clicks "Deepseek V4 Flash" under the Ollama group and the request silently routes to Xiaomi. Nobody reports this as a bug because the user can't tell — the model just behaves like a different model.

Your symptom report (label clutter) was the visible tip of a deeper iceberg (silent mis-routing). The proper fix has to be at the assignment site, not the dedup function.

What shipped

api/config.py:2078 now wraps auto_detected_models in copy.deepcopy() when assigning to a group, so each group gets its own independent dicts and dedup mutation cannot bleed across groups. The existing _deduplicate_model_ids() logic is unchanged.

Single-parenthetical disambiguation in labels is retained because the composer chip at static/index.html:441 shows the model label without the optgroup header context — Deepseek V4 Flash (Ollama) is more useful there than ambiguous Deepseek V4 Flash.

4 regression tests in tests/test_issue1511_dedup_shared_reference.py pin both the contract (independent dicts produce correct ids/labels) and the production path (get_models_grouped() actually emits independent dicts — checked via inspect.getsource() for the literal copy.deepcopy() call).

3925 → 3929 tests passing.

Why this matters for you

If you have multiple unconfigured auto-detected providers in your config, this release fixes silent model-routing. Pre-fix, selecting a duplicate model under any group routed the request to the alphabetical-first provider regardless of which group you clicked. Post-fix, requests route to the correct provider. You may notice your model selections behave differently after upgrading — that's the routing being correct, not a regression.

Thanks again for the report. The fix wouldn't have happened without your symptom description.