feat: add TOON output format (--format toon)

[davidfierro]

Pre-review checklist

• If my changes add or modify user-facing interface (commands, flags, output formats), I consulted maintainers and got sign-off beforehand (how). — Requesting design sign-off in this PR; see "Design sign-off requested" below.
• I've confirmed that instructions included in README.md are still correct after my changes in the codebase.
• I've added or updated unit tests to verify correctness of my new code. (11 TOON tests in tests/output/test_printing.py, round-trip via toon_format.decode)
• I've added or updated integration tests to verify correctness of my new code. — Happy to add once the design is signed off.
• Manually tested on macOS
• Manually tested on Windows
• I've confirmed my changes are up-to-date with the target branch. (rebased on latest main)
• I've described my changes in RELEASE-NOTES.md (see when and how).
• I've updated documentation if behavior changed. — No in-repo user docs mention output formats; help-text snapshots regenerated.

Changes description

Adds TOON (Token-Oriented Object Notation, spec) as a new structured output format: snow sql -q "..." --format toon (available on every command, like CSV/JSON).

Why TOON

TOON is a compact, lossless text format designed for LLM consumption — YAML-style indentation plus CSV-style tabular arrays, which is exactly the shape of snow sql query results. Numbers below are from the project's published benchmarks:

• Token efficiency: 39.9% fewer tokens than JSON on mixed datasets; up to −59.0% on fully tabular data (9,115 vs 22,245 tokens on time-series rows — the snow sql result shape). E-commerce −33.3%, nested config −31.9%.
• LLM comprehension is not worse than JSON: mean retrieval accuracy 76.4% (TOON) vs 75.0% (JSON) while using 39.9% fewer tokens; TOON beats JSON on 3 of 4 evaluated models (Claude Haiku 59.8 vs 57.4, GPT-5 Nano 90.9 vs 89.0, Grok-4 58.4 vs 56.5; Gemini 3 Flash essentially tied at 96.7 vs 97.1). Field-retrieval accuracy 99.6%.
• Use case: snow sql output consumed by agents/LLM pipelines — materially lower context cost with no loss of parseability (encode/decode round-trips are lossless; asserted in the unit tests).

Implementation

• New OutputFormat.TOON enum value and an is_structured property that consolidates the previously-triplicated "structured format" check (printing dispatch, intermediate-output muting, streamlit log streaming).
• TOON printer mirrors the CSV dispatch shape. Rows are pre-converted by _to_encodable_value to preserve data fidelity: Decimal as exact string (never float), bytes/bytearray as hex, non-finite floats as NaN/Infinity tokens (matching JSON output, since the toon encoder would collapse them to null), dates as ISO-8601, keys and string values passed through sanitize_for_terminal. Empty single-row results print null, matching JSON semantics.
• Related fix: several commands gated their structured return on output_format.is_json (stage diff, app diff, app validate, app version create, app release-channel list, apps init, helpers write-project-definition-schema). Under --format csv these emitted no output at all (console muted + None return). They now use is_structured, so CSV and TOON get structured results.
• Dependency: toon-format==0.9.0b1 (MIT, zero runtime dependencies, Python ≥3.8). Note: this pins a beta — the 0.1.0 release on PyPI is a stale placeholder; 0.9.0b1 is the spec-aligned implementation. Flagging explicitly for dependency review.

Design sign-off requested

Per adding-commands.md, a new output format needs maintainer sign-off — requesting it here:

1. Format value: --format toon (case-insensitive, like the existing values).
2. Lifecycle stage: proposed GA alongside CSV/JSON; if you prefer PrPr/PuPr, I'm happy to gate it behind a FeatureFlag (e.g. ENABLE_TOON_OUTPUT_FORMAT) in this PR.
3. Result mapping: collections → tabular TOON; objects/messages → key-value TOON; multiple results → one TOON document per result separated by a blank line (mirrors CSV); empty single result → null (mirrors JSON).

Proposed follow-ups (out of scope here)

• Consolidate the four parallel scalar-conversion tables in printing.py (CustomJSONEncoder, StreamingJSONEncoder, _write_csv_row, _to_encodable_value) into one shared converter.
• Sanitize CSV header keys (same terminal-safety gap, pre-existing).
• Fix the dead-code type guard in _stream_collection_as_csv (raises after the .keys() call that already crashes).