feat(vortex-file): gate the file writer behind editions

[joseph-isaacs]

Rationale for this change

Follow-up to the editions implementation (#8871): the edition declarations existed but nothing enforced them. This PR gates the file writer behind editions. The session is the source of truth: it stores the set of enabled editions alongside the declarations, and exposes the resolved set of all enabled encodings. The default session enables core2026.07.0 (EditionId::new("core", 2026, 7, 0)), so by default:

• the compressor only chooses schemes whose output encodings belong to the enabled editions, and
• the set that validates every encoding serialized to disk (failing the write on a violation) is that same edition's encoding set.

What changes are included in this PR?

• vortex-edition: EditionSession now stores the enabled editions next to the declarations: enable() (must be declared; at most one edition per family), enabled(), and enabled_encodings() — the resolved, sorted union of the enabled editions' encoding sets (None when nothing is enabled). The first-party declarations stay in the vortex facade; this crate remains types + session variable + test harness.
• vortex (facade): register_default_editions now also enables DEFAULT_WRITE_EDITIONS (core2026.07.0, plus the newest unstable draft when the unstable_encodings feature opts in) on the session. New tests: a golden pin of the frozen core2026.07.0 encoding set, per-edition harness validation, subset properties, and a check that every default compression scheme's produced encodings are covered by the default write editions.
• vortex-compressor: new required Scheme::produced_encodings(&self) -> Vec<ArrayId> trait method declaring the array encodings a scheme may emit (empty = only canonical arrays; cascaded children are covered by their own schemes). CascadingCompressor::with_enabled_editions_gating makes scheme selection resolve the session's enabled encodings from the execution context at compression time and skip schemes whose outputs are not all allowed; sessions with no enabled editions are not gated. Unit-tested with a synthetic edition/scheme.
• vortex-btrblocks: every scheme declares its produced encodings (including the env-gated Patched output of BitPacking/ALP). BtrBlocksCompressorBuilder gains gate_by_enabled_editions() (session gating pass-through) and retain_allowed_encodings() (static filtering for explicit allow sets).
• vortex-layout: FlatLayoutStrategy::with_enabled_editions_gating resolves the allowed encodings from the session's enabled editions at write_stream time — an encoding outside them fails the write via the existing normalize-with-Error validation — falling back to the explicit allowed_encodings set when the session enables no editions.
• vortex-file: the default WriteStrategyBuilder wires all of this up: flat writer gated by enabled editions with the static ALLOWED_ENCODINGS registry as fallback, and both the data and stats compressors gated at compression time. An explicit with_allow_encodings set overrides the session gate everywhere (unchanged behavior for vortex-jni / vortex-parquet-variant). The writer's deterministic pre-populated ArrayContext likewise comes from the session's enabled encodings. Experimental patches remain a runtime opt-out: Patched is only writable when the experimental environment variable is set, mirroring when compression produces it.

Checks run (after rebasing onto develop at #8871): cargo nextest run -p vortex-edition -p vortex-compressor -p vortex-btrblocks -p vortex-layout -p vortex-file -p vortex (411 passed; also with --all-features for file/btrblocks/compressor), cargo test --doc for the touched crates, cargo +nightly fmt --all, and cargo clippy --all-targets --all-features for the touched crates plus vortex-tensor, vortex-jni, and vortex-parquet-variant (clean). Workspace-wide clippy could not be run: vortex-duckdb's build script cannot download DuckDB sources in this sandbox.

What APIs are changed? Are there any user-facing changes?

• Scheme has a new required method produced_encodings() -> Vec<ArrayId> — a breaking change for third-party schemes, which must now declare their outputs.
• EditionSession::{enable, enabled, enabled_encodings}, CascadingCompressor::with_enabled_editions_gating, BtrBlocksCompressorBuilder::{gate_by_enabled_editions, retain_allowed_encodings}, FlatLayoutStrategy::with_enabled_editions_gating, and vortex::editions::DEFAULT_WRITE_EDITIONS are new.
• Behavior change for sessions built via the vortex facade: encodings outside the enabled editions (fastlanes.delta, vortex.onpair, vortex.zstd_buffers without the unstable_encodings feature, vortex.patched without the env opt-in) are no longer writable — writes containing them fail at the flat-layout validation instead of producing files without the read-forever guarantee. Sessions that enable no editions keep the previous static ALLOWED_ENCODINGS behavior.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CK2nCuXnyd2g3mNHQC8Lz2

[github-actions]

Polar Signals Profiling Results

Latest Run

Status	Commit	Job	Attempt	Link
🟢 Done	c84896c		1	Explore Profiling Data

Previous Runs (1)

Status	Commit	Job	Attempt	Link
🟢 Done	2da25df		1	Explore Profiling Data

---

Powered by Polar Signals Cloud

[github-actions]

Benchmarks: Vortex queries 📖

Verdict: No clear signal (low confidence)
Attributed Vortex impact: +9.7%
Engines: DataFusion No clear signal (+7.3%, environment too noisy confidence) · DuckDB Likely regression (+12.2%, high confidence)
Vortex (geomean): 1.092x ➖
Parquet (geomean): 1.042x ➖
Shifts: Parquet (control) +4.2% · Median polish +8.8%

How to read Verdict and Engines

• Verdict: Overall PR-level signal after subtracting baseline drift estimated from Parquet control rows. It can be Likely improvement, Likely regression, or No clear signal.
• Engines: Per-engine attribution. DataFusion is compared against DataFusion/Parquet controls; DuckDB is compared against DuckDB/Parquet controls. This answers whether each engine improved or regressed independently.
• Confidence: Based on directional consistency, share of rows above the noise floor, and control-run noise.

---

datafusion / vortex-file-compressed (1.127x ❌, 0↑ 2↓)

name	PR c84896c (ns)	base 94e593e (ns)	ratio (PR/base)
vortex_q00/datafusion:vortex-file-compressed 🚨	10735130	9541200	1.13
vortex_q01/datafusion:vortex-file-compressed 🚨	6741768	5974462	1.13

datafusion / parquet (1.050x ➖, 0↑ 0↓)

name	PR c84896c (ns)	base 94e593e (ns)	ratio (PR/base)
vortex_q00/datafusion:parquet	21479881	21214230	1.01
vortex_q01/datafusion:parquet	5161545	4736768	1.09

duckdb / vortex-file-compressed (1.160x ❌, 0↑ 2↓)

name	PR c84896c (ns)	base 94e593e (ns)	ratio (PR/base)
vortex_q00/duckdb:vortex-file-compressed 🚨	11416035	9885437	1.15
vortex_q01/duckdb:vortex-file-compressed 🚨	7028035	6028973	1.17

duckdb / parquet (1.034x ➖, 0↑ 0↓)

name	PR c84896c (ns)	base 94e593e (ns)	ratio (PR/base)
vortex_q00/duckdb:parquet	24060989	23250798	1.03
vortex_q01/duckdb:parquet	9689434	9373387	1.03

---

File Size Changes (1 files changed, +0.0% overall, 1↑ 0↓)

File	Scale	Format	Base	HEAD	Change	%
test.vortex	1.0	vortex-file-compressed	80.59 MB	80.59 MB	+248 B	+0.0%

Totals:

• vortex-file-compressed: 80.85 MB → 80.85 MB (+0.0%)