Noyalib external positioning

noyalib external positioning

Scope. Everything the project says about itself in public — READMEs, release notes, MIGRATION-FROM-*.md, COMPARISON.md, BENCHMARKS.md, blog posts, papers, tweets, conference talks. Load before touching any of that surface. For internal docs (USER-GUIDE, ARCHITECTURE, POLICIES) load noyalib-docs-and-writing. For unproven / speculative directions load noyalib-research-frontier.

Snapshot date: 2026-07-08. The ecosystem below drifts. Re-verify before shipping copy that names a peer.

Reference material (the per-peer ecosystem map, the reproducibility standards, the migration-audience playbook, and the pre-publication provenance checklist): see reference.md in this directory. This file keeps the ground-truth file list, the provable differentiators, and the claiming rules (project law).

---

1. Ground truth (read before making a claim)

• doc/COMPARISON.md — the ecosystem matrix. Cell states dated 2026-05 in-file; re-check crates.io + each repo before publication.
• doc/BENCHMARKS.md — headline speedups, methodology, hardware disclosure (Apple M4, aarch64-apple-darwin, Rust 1.95 stable, criterion --warm-up-time 2 --measurement-time 4, --release LTO=fat cgu=1 panic=abort). Reproducible via cargo bench --bench comparison.
• README.md — user-facing top-of-funnel claims; the "One-minute migration" table lists every peer we court.
• doc/MIGRATION-FROM-*.md — per-crate function tables (serde_yaml, serde_yaml_ng, serde_yml, serde-norway, serde-yaml-bw, serde-saphyr, yaml-serde, yaml-spanned).
• crates/noyalib/benches/comparison.rs — the benched peers today are serde_yaml_ng, yaml-rust2. serde_yml / libyml were dropped in v0.0.6 over RUSTSEC-2025-0067 + RUSTSEC-2025-0068 (doc/POLICIES.md §3, CHANGELOG under v0.0.6).
• crates/noyalib/tests/competitive_features.rs, competitive_features_full.rs, competitor_bugs.rs — parity tests + reproductions of upstream bugs (e.g. yaml-rust2#70 UTF-8 flow).
• crates/noyalib/src/de/config.rs — the current authoritative list of ParserConfig::max_* builder methods (do not repeat a hand-typed budget count — grep the source).

---

2. Ecosystem map (2026-07 — verify each row before publication)

Give every peer its honest claim to fame, then noyalib's honest differential — verifiable facts only, no disparagement. The migration wave noyalib courts is the serde_yaml (archived 2024) refugees; serde_yaml_ng (libyaml-backed) is the primary bench peer; saphyr/serde-saphyr and yaml-spanned are benched peers (saphyr is also an opt-in noyalib dep — not an adversary); serde_yml was retired from bench dev-deps in v0.0.6 over RUSTSEC-2025-0067/-0068. Full per-peer map with each migration guide in reference.md §R1.

Rule: before naming any peer in copy, cargo search <peer> + open the repo; a release in the last 90 days may invalidate a "unique to noyalib" line.

---

3. Provable differentiators (each with WHERE the proof lives)

Everything here is CI-enforced or bench-reproducible. If you cannot point at a file, remove the claim.

• Zero unsafe (workspace-wide). #![forbid(unsafe_code)] at every crate root (see crates/noyalib/src/lib.rs) plus unsafe_code = "forbid" in the [lints.rust] table of crates/noyalib/Cargo.toml. Enforcement is at the rustc / Cargo-lint layer, not a cargo geiger job — no such job exists in .github/workflows/. Proof: grep -R "unsafe" crates/ returns zero blocks in first-party code. Peers with C-FFI (serde_yaml_ng, serde_yml) fail this via unsafe-libyaml in the tree — verify via cargo tree.

• Spans + Spanned<T>. Every deserialized field can carry (line, column, byte_offset) and serialize transparently as T. COMPARISON.md marks this yes/no across the matrix; re-verify against yaml-spanned and saphyr before writing "unique among serde-YAML crates" — both have span support in some form. The current honest framing is: "spans with a serde-typed Spanned<T> wrapper that is transparent to the serde surface."

• Lossless CST editing (the moat). noyalib::cst::Document + set(path, value) — span-precise edits that preserve original formatting (comments, blank lines, indent style). No peer in COMPARISON.md ships this. Proof: the CST module, the recent span-fidelity commits (see recent CHANGELOG entries around alias span-through, block-collection value spans, keep-chomped scalars).

• DoS budget suite. ParserConfig exposes a suite of max_* builders. Do not hard-code a count in copy — grep source of truth: grep -c "pub fn max_" crates/noyalib/src/de/config.rs. Current builders include max_depth, max_document_length, max_alias_expansions, max_mapping_keys, max_sequence_length, max_events, max_nodes, max_total_scalar_bytes, max_documents, max_merge_keys, plus max_include_depth. README §"Configurable resource budgets" lists the default vs strict() values. Peer comparisons: COMPARISON.md marks serde_yml/serde_yaml_ng as "Basic", saphyr/rust-yaml as "Yes", yaml-rust2 as "No" — state peers' posture only if COMPARISON.md still says so.

• Official-suite conformance. Ground truth lives in the tests: crates/noyalib/tests/official_suite.rs sets SKIP_LIST = &[] (line 164) and asserts compliance >= 94.0 (line 312); yaml_compliance_report.rs mirrors the empty skip list. 351 wrapper .yaml files under tests/yaml-test-suite/ unpack to a larger number of case documents. Prose in README.md ("387/387, 19 skipped") and doc/BENCHMARKS.md ("406/406, 0 skipped") disagrees with the tests — that is in-tree documentation drift. Cite the tests, not the prose: "no skip list, ≥94 % compliance floor, empty SKIP_LIST asserted". Re-run before releasing any specific pass-count claim. The other libraries either don't track this publicly or apply lenience.

• Supply-chain posture. SLSA L3 provenance, sigstore keyless signing, cargo-deny / cargo-vet / cargo-semver-checks gates, REUSE.software 3.3 compliance, signed commits, OpenSSF Scorecard ~9/10 (was 6.5 before v0.0.6) — see doc/POLICIES.md §3. Every artifact carries a cosign keyless signature + SLSA L3 attestation. noyalib currently targets the OpenSSF Best Practices Badge for the v0.1 milestone (not yet earned — do not claim it).

---

4. Claiming rules (project law)

These are non-negotiable. Break one and the release notes have to be retracted.

1. No performance claim without a criterion run in the same PR. Hardware, toolchain, and flags disclosed exactly the way doc/BENCHMARKS.md opens: chip, target triple, Rust version, criterion flags, --release + LTO / codegen-units / panic settings. "Fast" / "high-performance" are fine as flavour text; any number needs a reproducible bench.

2. No "faster than X" without a runnable comparison. The comparison bench (benches/comparison.rs) must include X in the same PR that ships the copy. Third parties must be able to reproduce with a single cargo bench --bench comparison. Peers dropped from bench dev-deps (currently serde_yml, libyml) cannot appear in perf copy without being re-added and the RUSTSEC exposure re-evaluated.

3. No uniqueness claim without checking current peer versions. Peers evolve. "Only Rust YAML library that does X" ages fast. Re-verify against the last crates.io release of every peer named in COMPARISON.md before shipping the claim. If in doubt, downgrade to "the only [category] we know of" or "unique among the peers in COMPARISON.md as of 2026-07-08."

4. Unproven = labeled candidate/open. If it isn't measured or CI-enforced, the copy says "candidate", "planned", "roadmap", or "goal" — never a bare present-tense assertion. Load noyalib-research-frontier for the speculative bucket.

5. Security claims only for what CI enforces. "Zero unsafe" is fine because #![forbid(unsafe_code)] gates it. "DoS-safe" is fine because the budget suite is on by default and the strict mode is documented. "Memory-safe" without a Miri run in CI is not fine. "Constant-time" claims require a dudect / cachegrind measurement, which we do not currently ship.

6. No disparagement. State only verifiable facts about peers. The RUSTSEC-2025-0067/-0068 note about libyml is a fact; framing it as a value judgement about the maintainer is not. Ecosystem goodwill is a project asset.

---

5. Reproducibility standards

Every measured claim must be re-runnable by a third party from a fresh clone with one command, disclosing hardware, exact toolchain, build flags (--release LTO=fat cgu=1 panic=abort), and criterion knobs — the pattern doc/BENCHMARKS.md opens with. Reproducible builds are a 1.0 gate, not shipped — claim the goal, not the fact. Full standard (the disclosure checklist + the CST round-trip reproducibility rule) in reference.md §R2.

---

6. Migration-audience playbook

The migration wave is the primary v0.0.x audience; pitch order is drop-in shim → per-crate guide → upgrades, and the upgrades land in the order a serde_yaml refugee cares about (correctness → security → spans → CST editing → performance), because "will my code break?" is the first fear and "is it faster?" the last. Full playbook (shim mechanics, the eight MIGRATION-FROM-*.md guides, ordered rationale) in reference.md §R3.

---

7. When NOT to use this skill

• Writing the docs themselves (USER-GUIDE.md, ARCHITECTURE.md, POLICIES.md, per-module docstrings) → load noyalib-docs-and-writing.
• Research / speculative directions not yet measured or CI-enforced → load noyalib-research-frontier.
• Architecture invariants of the library itself → load noyalib-architecture-contract.
• Changing CI / release plumbing → load noyalib-ci-and-release.

---

8. Provenance / re-verification checklist

The pre-publication re-verify one-liners (COMPARISON.md header date, BENCHMARKS.md methodology, current benched peers, retired-peer rationale, live budget-builder count, peer freshness via cargo search) live in reference.md §R4. Run them before merging any copy that names a peer, benchmarks a peer, or makes a uniqueness claim.

Snapshot date on this skill: 2026-07-08. Anything that names a peer version, a number, or a "unique" claim is volatile — re-verify before publication.