docs: bring skills and crate READMEs back in line with main

crates/sparq-algos/README.md

@@ -63,7 +63,8 @@ let order = topological_sort(&g)?;                         // Err(CycleError) on
 - **Degree centrality** — In / Out / Total, raw counts or normalised to `[0, 1]`, plus a
   deterministic `top_k`.
 - **Extended centrality** — with the default-OFF `centrality-extended` feature, exact
-  unweighted Brandes betweenness and normalised harmonic closeness. Both treat each directed
+  unweighted Brandes betweenness, normalised harmonic closeness, and `core_number`
+  (k-core decomposition, O(V + E)). All three treat each directed
   edge as an undirected connection; reciprocal edges and self-loops do not multiply paths.
   Harmonic closeness assigns unreachable nodes a zero contribution, so disconnected graphs
   have finite scores.
@@ -73,7 +74,7 @@ let order = topological_sort(&g)?;                         // Err(CycleError) on
   strongly connected components, their count accessor, and a canonical topological sort.
   SCC ids are densified by ascending node index; topological-sort ties choose the smallest
   ready node and cycles, including self-loops, return `CycleError`; `is_acyclic` exposes the
-  same check as a boolean. [GPT-5.6] sq-awq7n.
+  same check as a boolean. sq-awq7n.
 - **Opt-in & lean** — consumes only sparq-core's public read API; the only dependencies
   are `sparq-core`, `oxrdf`, and `rustc-hash`. The heavier all-pairs algorithms are behind
   `centrality-extended`, and directed topology is behind `topology`; both features are OFF

crates/sparq-arrow/README.md

@@ -1,15 +1,15 @@
 # sparq-arrow
 
-> 🤖 **SPARQ agent** [OPUS-4.8] — opt-in **Apache Arrow** columnar export for sparq.
+> 🤖 **SPARQ agent** — opt-in **Apache Arrow** columnar export for sparq.
 > Projects a SPARQL `SELECT` result (`sparq_engine::QueryResult`) into an Arrow
 > `RecordBatch` so query results flow into the dataframe / analytics / ML ecosystem
 > (Polars, DuckDB, pandas) without a CSV round-trip. Issue #910, bead `sq-v78l4`.
 >
-> [GPT-5.6] Bead `sq-lsp7k.16` adds the checked inverse, `from_record_batch`.
-> [GPT-5.6] Bead `sq-lsp7k.21` adds Parquet byte serialization over that same schema.
-> [GPT-5.6] Beads `sq-r3cab` and `sq-ksxa2` add Arrow IPC bytes and schema-only readers.
-> [GPT-5.6] Bead `sq-kix7x` adds a metadata-only Parquet row-count reader.
-> [FABLE-5] Bead `sq-lsp7k` adds flattened CSV bytes (hand-rolled RFC 4180, zero new deps).
+> Bead `sq-lsp7k.16` adds the checked inverse, `from_record_batch`.
+> Bead `sq-lsp7k.21` adds Parquet byte serialization over that same schema.
+> Beads `sq-r3cab` and `sq-ksxa2` add Arrow IPC bytes and schema-only readers.
+> Bead `sq-kix7x` adds a metadata-only Parquet row-count reader.
+> Bead `sq-lsp7k` adds flattened CSV bytes (hand-rolled RFC 4180, zero new deps).
 
 **Opt-in / lean-core by construction.** This is a separate leaf crate, and Arrow
 import/export sits behind the `arrow` feature (OFF by default). The `arrow-*` dependency
@@ -39,8 +39,8 @@ assert_eq!(restored.rows, result.rows);
 # Ok::<(), Box<dyn std::error::Error>>(())
 ```
 
-`cargo add sparq-arrow --features arrow` (or, in this workspace,
-`cargo build -p sparq-arrow --features arrow`).
+`cargo add sparq-arrow --git https://github.com/sparq-org/sparq --features arrow`, with `sparq-core` and
+`sparq-engine` from the same git source so `QueryResult` types match (in-workspace: `cargo build -p sparq-arrow --features arrow`).
 
 To serialize the same term-struct batch as an in-memory Parquet file, enable the
 default-OFF `parquet` feature:
@@ -60,7 +60,7 @@ assert_eq!(restored.rows, result.rows);
 # Ok::<(), Box<dyn std::error::Error>>(())
 ```
 
-Use `cargo add sparq-arrow --features parquet` (or
+Use `cargo add sparq-arrow --git https://github.com/sparq-org/sparq --features parquet` (or
 `cargo build -p sparq-arrow --features parquet`); `parquet` implies `arrow`.
 
 For an Arrow IPC stream instead, enable the default-OFF `ipc` feature and call

crates/sparq-bench/README.md

@@ -1,4 +1,4 @@
-<!-- [OPUS-4.8] sq-4kr5: internal-stub README for a publish=false crate. -->
+<!-- sq-4kr5: internal-stub README for a publish=false crate. -->
 # sparq-bench
 
 The **benchmark + differential harness** for [sparq](../../README.md): it runs

crates/sparq-canon/README.md

@@ -10,7 +10,6 @@ signing, diffing, deduplication, and content-addressing.
 
 [RDF Dataset Canonicalization]: https://www.w3.org/TR/rdf-canon/
 
-> Model: Opus 4.8 (Fable unavailable — flag for re-review when Fable returns).
 > Surfaced from `sparq-zk::canon` per bead sq-0qip.
 
 ## 🚀 Quickstart
@@ -101,9 +100,9 @@ is **not** independent of RDFC-1.0 itself.
 
 ### Opt-in, single-sourced
 
-`publish = false`, and nothing in sparq's default dependency graph or the wasm
-artifact depends on this crate, so both are byte-identical with or without it and
-`sparq-core` stays lean. The RDFC-1.0 **algorithm** is the maintained zkp-ld
+`publish = false`; no published sparq crate nor the default wasm artifact depends on it
+(it builds by default only as a `sparq-bench` dependency; `sparq-wasm` pulls it behind
+its opt-in `canon` feature), so `sparq-core` stays lean. The RDFC-1.0 **algorithm** is the maintained zkp-ld
 [`rdf-canon`](https://crates.io/crates/rdf-canon) crate (oxrdf 0.2); this crate
 owns the single canonical-N-Quads-text bridge from sparq's oxrdf 0.3, so the
 bridge lives in exactly one place. `sparq-zk` depends on it (its `canon` module

crates/sparq-cli/README.md

@@ -46,11 +46,11 @@ cargo run --release -p sparq-cli -- query data.ttl turtle 'SELECT * WHERE { ?s ?
   `turtle`/`trig` straight to stdout (byte-identical, no whole-document `String`); a broken pipe
   (`dump … | head`) exits quietly instead of panicking.
 - **`to-hdt <file> <in-fmt> <out.hdt[.gz|.zst|.bz2]>`** *(opt-in `hdt-write` feature, which
-  implies `hdt`; [FABLE-5] sq-8ju74)* — export a loaded document (any ingestible format, HDT
+  implies `hdt`; sq-8ju74)* — export a loaded document (any ingestible format, HDT
   itself included) as a standard-layout **HDT v1.0 archive** via `sparq-hdt`'s direct in-memory
   encoder; the output container is chosen by the output extension. HDT holds a single default
   graph: named graphs are dropped **loudly** (a stderr warning with the dropped counts).
-- **JSON-LD I/O is default-on** ([OPUS-4.8] sq-oy1f.4, user-prioritised epic sq-oy1f) — the
+- **JSON-LD I/O is default-on** (sq-oy1f.4, user-prioritised epic sq-oy1f) — the
   default CLI **reads** a JSON-LD document (`<in-fmt>` ∈ `jsonld` / `json-ld` / `application/ld+json`;
   `@graph` named graphs are preserved as a dataset) AND **writes** one (the `jsonld*` out-formats
   above), out of the box. This is a deliberate maintainer-directed exception to sparq's
@@ -59,10 +59,10 @@ cargo run --release -p sparq-cli -- query data.ttl turtle 'SELECT * WHERE { ?s ?
   JSON-LD **parse + serialise + full 1.1 Compaction/Framing**; full conneg-conformance ratcheting
   is on the [sq-oy1f](https://github.com/sparq-org/sparq/issues/757) roadmap.
 - **`--reason <rdfs|owl-rl|n3>`** — opt-in forward-chaining materialization before query.
-- **`--reason datalog:<rules.dlog>`** *(opt-in `datalog`; [SONNET-4.6] sq-p4zci)* — run a **stratified Datalog** program
+- **`--reason datalog:<rules.dlog>`** *(opt-in `datalog`; sq-p4zci)* — run a **stratified Datalog** program
   over the parsed triples, then query the closure: negation as failure + `AGGREGATE`/`FILTER`, which monotone RDFS/OWL-RL
   cannot express. Loud exit-1 outside the fragment or on a `NOT`/`AGGREGATE` cycle; exit 2, no fall-back, when off.
-- **`classify <file> <format> [out.nt]` / `--reason el`** *(opt-in `el` feature; [OPUS-5] sq-2ch27)* —
+- **`classify <file> <format> [out.nt]` / `--reason el`** *(opt-in `el` feature; sq-2ch27)* —
   run the **OWL 2 EL** consequence-based classifier (`sparq-reason-el`, pulled with its `rbox` role
   automaton) and materialize the class-subsumption lattice as `rdfs:subClassOf` triples (plus the
   role-inclusion closure as `rdfs:subPropertyOf`) — complete for the **E1+E2 fragment** (CR1–CR6
@@ -75,27 +75,27 @@ cargo run --release -p sparq-cli -- query data.ttl turtle 'SELECT * WHERE { ?s ?
   `el`, not `--reason owl`, when you need the EL class hierarchy. Honest incompleteness (skipped
   axioms, a non-regular RBox, unsatisfiable classes) is reported on stderr, never swallowed;
   without the feature `--reason el` exits 2 naming it rather than silently downgrading to RL.
-- **`tabular <csv[.gz|.zst|.bz2]> …`** *(opt-in `tabular` feature; [FABLE-5] sq-lsp7k.8)* —
+- **`tabular <csv[.gz|.zst|.bz2]> …`** *(opt-in `tabular` feature; sq-lsp7k.8)* —
   **materializing tabular→RDF import**: stream CSV rows through a direct mapping (subject IRI
   template `{col}`/`{_row}` via `--template`, per-column predicates, `xsd` datatype inference,
   per-row `rdf:type`) or an **R2RML mapping** (`--mapping <r2rml.ttl>`, CSV logical tables bound
   by `rr:tableName` = file stem) into a loaded graph (`--query` runs SPARQL in the same shot) or
   an N-Triples stream (`--out out.nt[.gz|.zst]`). Streaming end-to-end (no whole-file buffering).
-  [OPUS-5] (sq-u1z86) adds cross-CSV **joins** (`rr:parentTriplesMap` + `rr:joinCondition`, run as
+  (sq-u1z86) adds cross-CSV **joins** (`rr:parentTriplesMap` + `rr:joinCondition`, run as
   a keyed hash join over a one-pass parent index), **named-graph** output (`rr:graphMap`/`rr:graph`
   → N-Quads + a dataset load, so `GRAPH ?g { … }` works) and `--row-provenance`
   (`prov:wasDerivedFrom` the source row). SQL-connection R2RML stays out of scope (`rr:sqlQuery`,
   `rr:sqlVersion`, `rr:inverseExpression` fail loudly — sparq materializes, it does not virtualize).
-- **`terse <query | ->`** *(opt-in `terse` feature; [OPUS-4.8] sq-vczh2)* — transpile a terse query
+- **`terse <query | ->`** *(opt-in `terse` feature; sq-vczh2)* — transpile a terse query
   (the `K:<name>` keyword layer over canonical SPARQL) into the **canonical SPARQL** it expands to,
   printing the verifiable JSON `{ canonical_sparql, keywords, resolutions, warnings, legendVersion }`
   (the same contract the server's `POST /terse/transpile` returns). It never executes the query —
   pipe `canonical_sparql` into `query`. Loud-fails (exit 2) on an unknown keyword or a `V(...)`
-  construct rather than guessing. Lean build — depends only on `spargebra`. `--features terse`.
+  construct rather than guessing. Lean build — depends only on `spargebra`. `--features terse`. (`terse` and `el` reach unpublished crates, so the crates.io package omits both; build from git.)
 - **Engine features the default CLI build lights** — the CLI's `sparq-engine` dependency enables
   `dp-planner` (DPccp cost-optimal join ordering, sq-7d3dj.30.5) and `algebra-rewrite` (the
   result-equivalent pre-execution rewrite of #1735 — `FILTER(?v = <iri>)` constant folding +
-  `!bound` anti-join; [FABLE-5] sq-7d3dj.30.13), so the shipped binary and every canonical
+  `!bound` anti-join; sq-7d3dj.30.13), so the shipped binary and every canonical
   benchmark run the same plans. The engine **library** defaults stay lean (both OFF there).
 - **Transparent decompression** — `.gz` / `.bz2` / `.zst` inputs detected by content.
   The gzip path defaults to the pure-Rust `miniz_oxide` backend; the opt-in, native-only
@@ -108,7 +108,7 @@ cargo run --release -p sparq-cli -- query data.ttl turtle 'SELECT * WHERE { ?s ?
 
 - **How-to** — [`skills/cli/SKILL.md`](../../skills/cli/SKILL.md) (full subcommand reference)
   and [`skills/inference/SKILL.md`](../../skills/inference/SKILL.md) (reasoning).
-- **API reference** — run `cargo run -p sparq-cli -- --help`; rustdoc at
+- **API reference** — run `cargo run -p sparq-cli` with no arguments for the usage block; rustdoc at
   [docs.rs/sparq-cli](https://docs.rs/sparq-cli).
 - **Design** — [`research/ARCHITECTURE.md`](../../research/ARCHITECTURE.md).
 - **Performance** — see the [benchmarks dashboard](https://sparq.jeswr.org/dev/bench);

crates/sparq-conformance/README.md

@@ -1,29 +1,29 @@
-<!-- [OPUS-4.8] sq-4kr5: internal-stub README for a publish=false crate. -->
+<!-- sq-4kr5: internal-stub README for a publish=false crate. -->
 # sparq-conformance
 
-The **W3C conformance harness** for [sparq](../../README.md): it runs the official
-test suites against the engine and reports a per-suite pass/fail/skip scoreboard,
-gated in CI by pass-count ratchets. Three binaries share manifest-walking /
-result-comparison machinery: `sparq-conformance` (W3C SPARQL query/update/syntax),
-`sparq-inference-conformance` (RDF Semantics, OWL 2 RL, N3, entailment regimes via
-`sparq-reason`), and `sparq-conformance-scoreboard` (a consolidated index of every
-ratchet — SPARQL, inference, SHACL, GeoSPARQL, Solid WAC + ACP, JSON-LD 1.1,
-SolidLab ODRL — plus eight `sparq extension` rows, HONESTLY labelled NOT standards
-claims and tallied separately). Floors are MEASURED and guarded textually; the
+The **W3C conformance harness** for [sparq](../../README.md): it runs the official test suites against
+the engine and reports a per-suite pass/fail/skip scoreboard, gated by pass-count ratchets (the W3C
+SPARQL one per PR in `ci-fast.yml`, the rest nightly / on dispatch in `ci.yml`). Three binaries share
+manifest-walking / result-comparison machinery: `sparq-conformance` (W3C SPARQL query/update/syntax),
+`sparq-inference-conformance` (RDF Semantics, OWL 2 RL, N3, entailment regimes via `sparq-reason`), and
+`sparq-conformance-scoreboard` (a consolidated index of every ratchet — SPARQL, RDF syntax, inference,
+RIF, SHACL, GeoSPARQL, Solid WAC + ACP, JSON-LD 1.1, SolidLab ODRL — plus eleven `sparq extension` rows,
+HONESTLY labelled NOT standards claims and tallied separately). Plus: `served-conformance-report`
+(served-surface SPARQL Protocol JSON report) and `sparq-notation3tests` (the ADVISORY community
+notation3tests runner, `notation3tests.yml`). Floors are MEASURED and guarded textually; the
 `scoreboard` rustdoc has the full per-lane provenance and divergence sets.
 
-The registry also has a **machine-readable export** (sq-gum8.14):
-`scoreboard::scoreboard_json()` renders the same rows + floors as deterministic JSON,
-committed as `bench/conformance-scoreboard.generated.json` and drift-guarded by
-`tests/scoreboard_export.rs` — so paper-evidence bindings can reference suite rows /
-floors by json-pointer without the mirror silently drifting. Several crate-local
-`cargo test` lanes sit behind **opt-in features** (OFF by default) — `jsonld-suite`,
-`service`, `http-protocol`, `federation-descriptors`, and the inference/geo/syntax
+The registry also has a **machine-readable export** (sq-gum8.14): `scoreboard::scoreboard_json()`
+renders the same rows + floors as deterministic JSON, committed as
+`bench/conformance-scoreboard.generated.json` and drift-guarded by `tests/scoreboard_export.rs` — so
+paper-evidence bindings can reference suite rows / floors by json-pointer without the mirror silently
+drifting. Several crate-local `cargo test` lanes sit behind **opt-in features** (OFF by default) —
+`jsonld-suite`, `service`, `http-protocol`, `federation-descriptors`, and the inference/geo/syntax
 lanes; the `scoreboard` rustdoc documents each lane's scope, floor and divergences.
 
 > **Internal dev-only harness — not published** (`publish = false`). Test data is
-> fetched by `scripts/fetch-conformance.sh` + the sibling `fetch-jsonld*` /
-> `fetch-odrl-suite.sh` scripts. Contributing: [`AGENTS.md`](../../AGENTS.md).
+> fetched by `scripts/fetch-conformance.sh` + the sibling `fetch-inference-suites.sh`
+> / `fetch-jsonld*` / `fetch-odrl-suite.sh` scripts. Contributing: [`AGENTS.md`](../../AGENTS.md).
 
 ## License
 

crates/sparq-core/README.md

@@ -50,7 +50,7 @@ assert_eq!(count, 1);
 - **Incremental updates** — start from `Graph::new()` / `Graph::default()` (an empty graph) and
   `insert_triple(s, p, o)` / `remove_triple(s, p, o)` a single triple from `oxrdf` terms, or apply
   a whole batch with `apply_delta` — in place, with an optional write-ahead log.
-- **Experimental deletion projections** — [GPT-6 Astra] enable the default-off
+- **Experimental deletion projections** — enable the default-off
   `overlay-deleted-projections` Cargo feature on `sparq-core` to cache sorted
   tombstone projections for repeated range counts. The default uses the original
   linear deletion scan and carries no deletion projection slots. Each requested
@@ -69,7 +69,7 @@ assert_eq!(count, 1);
   point lookups for an id that is ABSENT from the permutation (result-equivalent, never
   serialised; for a PRESENT id the zone map already narrows the candidate window to about one
   block, so there is little left to skip).
-  [GPT-6] Persisted predicate statistics have deterministic record order across builds and
+  Persisted predicate statistics have deterministic record order across builds and
   re-saves; older unordered statistics files remain readable.
 - **Compressed-seek column codecs (prototype)** — the opt-in `elias-fano` feature adds
   Elias-Fano and Partitioned-Elias-Fano codecs whose `next_geq(target)` answers a successor

crates/sparq-crdt/README.md

@@ -1,6 +1,6 @@
 # sparq-crdt
 
-<!-- [FABLE-5] sq-tag1q.7.2: internal-stub README for a publish = false crate. -->
+<!-- sq-tag1q.7.2: internal-stub README for a publish = false crate. -->
 
 Opt-in SPARQL-CRDT replication primitives for sparq (epic `sq-tag1q.7`): the
 bounded **canonical delta-envelope codec**, **causal summaries** (version-vector

crates/sparq-difftest/README.md

@@ -1,4 +1,4 @@
-<!-- [OPUS-4.8] sq-qcnn.4: internal-stub README for a publish=false crate. -->
+<!-- sq-qcnn.4: internal-stub README for a publish=false crate. -->
 # sparq-difftest
 
 The **engine-independent value-normalisation** library for [sparq](../../README.md)'s
@@ -11,7 +11,8 @@ equality (simple-literal ≡ `xsd:string`, case-insensitive language tags);
 arbitrary-precision integer/decimal value equality and `xsd:double` `INF`/`NaN`;
 `dateTime`/`date`/`duration` comparison with timezone normalisation and the ±14h
 indeterminate window; a SPARQL-Results-JSON reader; the multiset + `ORDER BY`
-equivalence-class comparators; and RDFC-1.0 canonical labelling for blank-node results,
+equivalence-class comparators; an admission-decision (allow/deny set) comparator with a
+fail-closed-asymmetry check (`decision`); and RDFC-1.0 canonical labelling for blank-node results,
 which are comparable only up to a bijection. Full API detail is in the rustdoc.
 
 **Load-bearing constraint:** it depends on **no sparq crate** — only third-party exact

crates/sparq-e2ee-ng/README.md

@@ -1,4 +1,4 @@
-<!-- [OPUS-4.8] sq-tag1q.9: internal-stub README for a publish=false crate; full posture + public-API surface live in skills/e2ee-ng/SKILL.md and the design record. -->
+<!-- sq-tag1q.9: internal-stub README for a publish=false crate; full posture + public-API surface live in skills/e2ee-ng/SKILL.md and the design record. -->
 # sparq-e2ee-ng
 
 Opt-in **E2EE-NG profile primitives** for sparq: the *capability / envelope /

crates/sparq-engine-serialize/README.md

@@ -5,11 +5,10 @@ The RDF **writer matrix** for [`sparq-engine`] — Turtle, TriG, N-Quads, and JS
 
 > **Internal, unstable crate.** This is seam 1 of the staged `sparq-engine` facade split
 > (RFC `research/engine-split-rfc.md` §4 Option A / §7 Phase A1, bead `sq-6vshe.4`). It is
-> `publish = false` and has **no stability guarantee of its own**. Depend on the
+> published to crates.io only to satisfy `sparq-engine`'s dependency closure and has
+> **no stability guarantee of its own**. Depend on the
 > re-export **`sparq_engine::serialize`** — its public API and the `serialize-rdf` /
 > `streaming-serialization` feature names are unchanged by the split.
->
-> Model: Opus 4.8 ([OPUS-4.8], Fable unavailable). Flag for re-review when Fable returns.
 
 [`sparq-engine`]: ../sparq-engine
 
@@ -48,8 +47,8 @@ assert!(jsonld.starts_with('['));
   builds of `sparq-engine` are byte-identical to before the split.
 - **`streaming-serialization`** *(feature, implies `serialize-rdf`)* — adds
   `write_turtle_streaming` / `write_trig_streaming` that render one subject block at a
-  time into a `std::io::Write`, for chunked CONSTRUCT/DESCRIBE responses without
-  materialising the whole rendered string. Byte-identical to the buffered output.
+  time into a `std::io::Write` without materialising the whole rendered string (what
+  `sparq-cli dump … turtle|trig` uses under its `streaming-serialization` feature). Byte-identical to the buffered output.
 
 ### Opt-in by construction