Add MiniMax summarizer provider

[octo-patch]

Reason: Make MiniMax directly configurable for trace summarization across supported regions and compatible protocols.

Changes:

• Add a MiniMax HTTP summarizer with current model selection, regional routing, and model-specific thinking behavior.
• Add persisted settings, connection testing, and onboarding controls for the new backend.
• Cover every region and protocol endpoint combination, response parsing, registry exposure, and config coercion.

Checks:

• git diff --check HEAD^
• PYTHONPATH=backend python3 backend/test_minimax_summarizer.py
• frontend/node_modules/.bin/eslint --config frontend/eslint.config.mjs frontend/src/app/settings/page.tsx frontend/src/components/summarizer/BackendPicker.tsx frontend/src/components/summarizer/OnboardingModal.tsx frontend/src/lib/agents.ts frontend/src/lib/summarizer.ts --rule 'react-hooks/set-state-in-effect: off'
• npm --prefix frontend run build

The repository-wide npm --prefix frontend run lint check still reports existing lint failures outside this change.

[VasiHemanth]

I've reverted the dedicated adapter on this branch and replaced it with endpoint presets on the existing openai_compat backend. Net diff vs main is now 3 files, +54/-4, instead of 10 files and +668/-19.

Why

MiniMax speaks the OpenAI chat-completions API, so it already worked through openai_compat before this PR. The only friction was knowing the base URL and the model id, which a preset solves.

Three things the dedicated adapter got wrong that the existing one already handles:

1. <think> blocks leak. On the OpenAI route with reasoning_split unset (the default), M3 inlines its reasoning into content wrapped in <think> tags. openai_compat strips those (_THINK_RE, backend/summarizers/openai_compat.py:53). _extract_openai_text in minimax.py did not, so the full chain-of-thought would reach parse_narrative, which expects JSON.

2. Hardcoded model allowlist. SUPPORTED_MODELS = ("MiniMax-M3", "MiniMax-M2.7") plus save_config silently rewriting model back to DEFAULT_MODEL when it isn't in that tuple (asserted on purpose in test_minimax_config_is_coerced_and_persisted). The next MiniMax model would be unusable until we cut a release. The free-text model field has no such gate.

3. Unrelated cross-backend changes. ModelDropdown's loading ? to loading || (!options && !error) ? changes rendering for ollama and codex too, and moving get_summarizer(...) inside the try in make_summary turns constructor-raised SummarizerError (such as openai_compat's SSRF scheme guard) into a soft gen_error for every backend. Neither is mentioned in the description.

Separately, every test in test_minimax_summarizer.py mocked _post, so the endpoint-matrix test asserted that the code builds the URL the code intends to build. Nothing validated the {"type": "adaptive"} thinking shape against a real response.

What's on the branch now

• ENDPOINT_PRESETS in frontend/src/lib/summarizer.ts: MiniMax, MiniMax (China), local llama.cpp. Each entry is an endpoint plus a starting model id.
• Preset buttons above the Endpoint field in the existing OpenAICompatForm. Clicking one fills endpoint and model; the API key and sampling knobs stay the user's.

Adding the next provider is one array entry.

Verified

• Unauthenticated POST through OpenAICompatSummarizer to https://api.minimax.io/v1/chat/completions returns MiniMax's 401 body, and errors.classify() buckets it as auth with the right hint.
• <think> inlining confirmed against MiniMax's OpenAI SDK docs.
• npx tsc --noEmit clean, UPDATE.json parses.

Needs a human check before merge

I don't have a MiniMax API key, so the authenticated end-to-end run hasn't happened. Please verify:

1. Settings, Summaries, pick "OpenAI-compatible server", click the MiniMax preset.
2. Paste your key, hit Test connection. Expect a short reply, not an error card.
3. Open a session and generate a summary. Confirm the narrative renders and no <think> text appears in it.

If step 3 shows reasoning text, that's a _THINK_RE gap worth fixing in openai_compat for all reasoning models, not a reason to bring the adapter back.

[VasiHemanth]

Follow-up on the local-server case.

The openai_compat adapter itself is untouched by this PR, so llama.cpp works exactly as before: the endpoint field still defaults to http://localhost:8080/v1, the non-OpenAI sampling extras (top_k, min_p, repetition_penalty) still go out, and the 400 retry still drops them for strict servers.

One bug in the preset buttons though, now fixed in 81595e7: a preset with an empty model id left the previous one in place, so clicking MiniMax and then the local preset would POST model: "MiniMax-M3" to localhost:8080. The click now reassigns unconditionally, and a null model falls through to self._model or "default" in openai_compat.py:132, which is what a server that serves whatever is loaded expects.

Add to the verification pass: after testing MiniMax, click the local preset and confirm the model field clears rather than keeping the MiniMax id.