feat(examples): make canonical SIE examples classroom-ready

[svonava]

Supports sie-internal#1963 by improving canonical SIE examples instead of creating course-specific forks.

Summary

The same examples now support a short classroom path while remaining useful to every SIE user. Connection changes are configuration only: local SIE, a self-deployed gateway, and managed SIE all use the public SDK.

Changes

Quickstart

• Add a 22-line examples/quickstart/quickstart.py: one SIEClient and one encode call, with no custom HTTP or JSON plumbing.
• Add a focused QuickStartTest and concise local/self-deployed/managed setup.
• Update the existing notebook to read SIE_CLUSTER_URL and optional SIE_API_KEY, starting a local server only when no remote endpoint is configured.

Retrieval ablation

• Add a fictional 12-document/6-query fixture.
• Add benchmark_ablation.py --sample, reusing the canonical NDCG@10, MRR@10, and Recall@10 evaluator.
• Isolate sample index, cache, namespace, and results from full benchmark runs.

Vision document RAG

• Add a deterministic, hash-verified three-image sample corpus sourced from the existing document-OCR fixtures.
• Keep it under data/sample/ without changing the primary page manifest.
• Add --pages-manifest so the canonical ingest pipeline can consume either corpus.

Contract review

• Commit the lockfile and document a frozen, offline synthetic-data quickstart before model-backed review.

There is no ABI/course-only package, manifest, schema, model-budget placeholder, or duplicated example implementation.

Validation

• Quickstart: 1 unit test passed.
• Retrieval sample: 3 tests passed, including canonical evaluator reuse and isolated dry-run state.
• Vision sample: 4 tests passed, including hash verification, deterministic replacement, stale-file cleanup, and explicit-manifest ingest.
• Contract review: uv sync --frozen, uv run make-sample, and uv run review --list passed.
• Ruff focused checks and formatting passed.
• Notebook JSON and git diff --check passed.

Live managed execution remains a release/acceptance step once the managed endpoint, active account, API key, and selected catalog models are available; this PR does not claim that validation has happened.

[svonava]

@coderabbitai review