[ai/azure-ai-projects] Eval sample fixes: expand finetuning seed and make simulation dataset always-fresh

[aprilk-ms]

Three small fixes uncovered while bug-bashing the public 2.2.0 samples against a live Foundry project.

1. Unicode print crashes on Windows cp1252 (16 sample files)

The "completed / failed" summary print statements at the end of 16 evaluation samples used U+2713 ✓, U+2717 ✗, U+2705 ✅, U+274C ❌. On a fresh Windows shell (stdout = cp1252, no PYTHONIOENCODING=utf-8), Python's stdio writer crashes with UnicodeEncodeError and the sample exits non-zero after the run already succeeded on the service. Three of my multi-turn samples failed this way before I forced UTF-8 in my shell.

Replaced all 36 occurrences with ASCII [OK] / [FAIL] tokens — pure mechanical change.

2. Finetuning seed document was too small (sample_dataset_generation_job_simpleqna_for_finetuning.py)

The embedded 1.2 KB SEED_REFERENCE_DOCUMENT passes the eval scenario but the SUPERVISED_FINETUNING scenario rejects it with:

File content lacks sufficient context to generate quality questions.

Confirmed empirically: an identical script with a ~10 KB version of the same widgets/gizmos/sprockets reference doc generates the requested 15 QnA pairs cleanly (96 s run). Expanded the seed to that ~10 KB version (same domain, much richer prose) and updated the surrounding comment to explain the requirement.

3. Simulation sample silently fell back to a stale cached dataset (sample_multiturn_conversation_simulation.py)

The previous code uploaded simulation-scenarios:v1, and on the inevitable "dataset already exists" failure for the second run it swallowed the exception and pointed the eval run at the cached server-side dataset URI. That made the sample non-reproducible — anyone editing the local JSONL would never see their changes, and the Expected: {3 * 2} print became wrong as soon as anyone uploaded a different-sized scenario file from a prior session (in my bug-bash session it printed Expected: 6 but actually ran 10 conversations).

Now the dataset name is suffixed with a per-run id (same pattern as the datagen samples) so every invocation uploads fresh data, and the "expected" count is computed from the actual JSONL row count.

Verification

• All 17 modified files py_compile clean.
• Live end-to-end run of sample_multiturn_conversation_evaluation.py on a default Windows shell now prints [OK] Evaluation run completed successfully! and exits 0 (previously: UnicodeEncodeError).
• All 14 attempted samples passed in my bug-bash with these fixes (rerun matrix in commit description).

Follow-up (not in this PR)

A few of the trace-based samples (sample_dataset_generation_job_traces_for_*, sample_multiturn_trace_evaluation_*) require pre-existing agent traces in the project to be runnable. I will open a separate issue/PR to make those self-contained by creating the agent and seeding traces inline.

[aprilk-ms]

Split into three focused PRs (rebased onto main) for easier review:

• [ai/azure-ai-projects] fix evaluation sample print crashes on cp1252 consoles #47253 — cp1252 print crash fix (16 evaluation samples, mechanical)
• [ai/azure-ai-projects] expand finetuning seed conversations so simpleqna sample succeeds #47254 — finetuning seed expansion (one sample, substantive)
• [ai/azure-ai-projects] make multiturn simulation sample's dataset always-fresh #47255 — multiturn simulation always-fresh dataset (one sample, substantive)