Fix selected configured model reporting in headless output and session metadata

[Brian Krabach (bkrabach)]

Problem

Actual headless CLI qualification with OpenAI SDK 3.22.1 found that selecting a non-first mounted provider sent the correct request but reported the first provider's model in JSON and saved session metadata. The initial --provider openai --model gpt-4.1 fixture reached gpt-4.1 while reporting vllm/meta-llama/Llama-3-8B. Four new Sol/Astra JSON/json-trace tests reproduced the same defect in both output and real SessionStore reloads before repair.

Bounded repair

• Touch only amplifier_app_cli/main.py and tests/test_headless_session_persistence.py.
• Display the configured conversation model using the existing conversation pin, then streaming-loop priority ordering with stable insertion-order ties. Read get_info().defaults['model'], with legacy attribute fallback. Never borrow a model from an unselected mount.
• Freeze one label for JSON and persistence before execution, inside the cleanup boundary.
• Add model_source='configured_default' or 'unknown'. This is configured-default reporting, not a claim of observed per-call routing, account identity, or the server's underlying model deployment.
• Best-effort reporting failure returns unknown; it must not discard a completed response or bypass cleanup.
• Preserve provider mount order, settings, routing, session ownership and canonical transcript content. No dependency change or new observer in production.

Trade-off: a scalar configured-default label does not summarize multi-model role routing. This PR makes its provenance explicit rather than adding unscoped event inference or changing selection.

Verification

Exact candidate 2b64f76; existing task-owned Linux-aarch64 DTU; Python 3.13.15, Core 2.0.1, Foundation 89575c3482e3e8afe5a03df72e723cf815fa1f6c, SDK 3.22.1. Normal package resolution and non-editable CLI wheel.

• Four baseline regression failures: Sol/Astra x JSON/json-trace reported the first mounted vLLM model, despite a selected OpenAI provider.
• Final default suite: 2478 passed, 4 skipped, 13 deselected, 1 xfailed, exit 0, 26.94 seconds. Skips: real macOS kqueue, two missing hooks-logging production-path checks, real tmux. Deselected tests are existing PTY integration tests.
• New priority/config/default/stable-tie, missing model, pin, and broken pin/model-property tests. Broken lookup cases prove successful response persistence and cleanup awaited once.
• Independent bounded source review's cleanup finding repaired and re-reviewed: no remaining critical source issue. Review is not GitHub approval.
• amplifier-tester's frozen CLI qualification verified 454 archive files and all 101 installed Python files against the candidate; installed main.py SHA256 6e5f3f0fba88f3ca3e455b2fa78a0075446e0c5dd15125990a1b457708e1d32b.
• Four actual CLI controlled-TCP cases: gpt-6.1-sol SSE, gpt-6-astra SSE, openai/gpt-oss-120b JSON, openai/gpt-oss-120b SSE, all with documented reasoning effort high. All execute the real read_file tool, return unpredictable fixture contents with invocation call_id distinct from output item id on the second HTTP request, and then receive the final success marker.
• Correct selected instance/model appears in both JSON and actual saved metadata for every case, with model_source='configured_default'. Sol/Astra omit unsupported sampling fields. SDK 3.22.1 remains installed before/after every CLI preparation.
• GPT-OSS uses production Harmony accounting, no accounting mocks: zero server usage is replaced with validated nonzero computed input/output usage, consistent with canonical provider events.
• 56 controlled-server assertions pass; 49 installed packages compatible. No shared-venv guard bypass; original DTU settings restored and test servers stopped.
• Root independently reread all case receipts, SDK snapshots and server assertions, then reran Sol SSE and GPT-OSS SSE against the same exact frozen CLI/provider bytes. Both exited0 with real tool round-trips, correct JSON/saved metadata and validated production Harmony accounting. Independent receipt phase3-20261001-010143-43a1f3 retained separately from the specialist's original result.
• GitHub CI36799174201 completed success on exact2b64f760: all eight jobs pass, including Python3.11/3.12 on Linux/macOS/Windows and real PTY integration on Linux/macOS. This supplements, rather than relabels, the DTU's explicitly deselected PTY tests.

The coupled vLLM SDK repair is microsoft/amplifier-module-provider-vllm#46 at 86d7d8ebcff4ba95427aa60bc0880d4632962b0e. Its own 333-test four-SDK matrix and all eight CI jobs are green. Neither PR is merged or installed on the live host by this qualification.

Limits: controlled local server, not real hosted model inference or production vLLM deployment; headless single-shot paths, not full interactive UX. An archive install's version banner falls back to 0.1.0 without .git; package metadata remains 0.1.1 and exact source identity is established by hashes, not the banner. Preserve initial failing receipts rather than relabeling them.