Registry pages: capture enumerable sources completely, one cited row per entry

[MarkusNeusinger]

What

Sources that are (or contain) a uniform enumeration of like entries — machine inventories, error-code catalogs, customer rosters, product/price lists — used to come out as a representative subset with the tail compressed into one footnote (live example: the kelvarra showcase's organizations/instrument-works.md presses 5 "Sundries" products into a single [^s9]). The rules actively demanded that compression: the granularity floor prefers fewer dense pages, "essence, not structure" discards enumerable data, and meeting-minutes.md routes perennial reference tables to "ordinary facts" with no aggregate destination.

This PR adds the Registry page kind: one page per collection under the new wiki/registries/ folder, one bullet per entry with the entry's source-verbatim key in bold and its own citation at the finest available locator — and promotion: an entry that accumulates several independent cited facts (the unchanged granularity floor) moves to its own page of its own kind, its registry row becoming key + link + one-line gloss.

How

Almost all of it is rules text (the runtime program):

• New citadel/rules/genres/registry.md — applies-when test ("entries share one attribute shape, and a reader would ask is X on the list?"), completeness-over-compression (WRONG→RIGHT built on the Sundries defect), row shape + key identity + search-before-append, spreadsheet sheet-marker rule, promotion, per-row reconcile/Change-Log behavior, range-split (never topic-split) guidance, and an explicit negative list (Open Points, chat/event streams, code symbols, bibliographies, abbreviations).
• schema.md — Registry row in the type table, decision-procedure step 7 (before Concept, after Object — so a single rich machine still matches Object first, which is the promotion semantics), and a compact ## Registries format contract.
• core.md — two carve-outs: the floor limits pages, not rows; data rows are not code structure.
• meeting-minutes.md — perennial reference tables now route to the registry brief; tasks/reconcile.md keeps registry row lists as genre treatment; rules README.md indexed.

Two small code touches:

• okf.py — "registry": "registries" in folder_for_type (same commit as the rules text: the misc-fallback plus curate's resort detector would otherwise relocate every registry page each run).
• curate.py — the page_length_hard guidance is now type-aware: an oversized registry is told to split by entry range/subclass into sibling registries, never along topics (the detector itself is unchanged).

Everything else (per-folder index, viewer, type:registry search, MCP resources, curate resort) is data-driven and picks the new type up automatically.

New graded corpus: werkhof — and its committed showcase run

Per the project's "every capability is graded" pattern: corpora/werkhof/ (fictional maintenance firm, Werkhof Anlagenservice Brandt) — a 28-machine register (>200 lines ⇒ per-row lines A-B locators required), a 20-code fault catalogue (near-miss E-142 vs E-412, deprecated E-155→E-310), a 15-customer CSV (the data-not-code boundary), and a service report driving promotion, a dated HX-201 status supersession, attributed opinions, and a done-able open-items list (Open Points, never registry rows). Hidden answer key at .claude/skills/verify-corpus/werkhof/ground-truth.md; SKILL.md/CLAUDE.md/CI corpus floors updated.

The committed showcase wiki was generated by a real sonnet Mode A run under the new rules and graded against the full answer key — PASS on every guarantee (run block in docs/verify-corpus-backlog.md): three type: Registry pages carrying 28/28 machines, 20/20 fault codes, 15/15 customers, one citation per row with offline-verifying locators (lint locator issues: 0), promotion to objects//organizations/ with rows reduced to key + link + gloss, the status supersession as a dated trace, open items as ## Open Points, and mid-list keys (KP-011, E-420, K-009) retrievable top-1 through citadel search. check and lint both exit 0. The kelvarra Sundries defect is recorded and resolved as VCB-005 in the ledger.

Known limitation (documented in CHANGELOG)

A list source over CITADEL_MAX_SOURCE_CHARS (~300k chars) chunks through segment files whose line numbers rebase, so per-row lines A-B locators of such oversized lists are not offline-verifiable yet. v1 registries (tens to a few hundred entries) sit far below the threshold; the future fix is the _line_windows true-offset approach already used for audio/PDF.

Tests

• tests/test_okf.py — Registry routing (+ the misc-fallback pins stay).
• tests/test_curate.py — Registry in registries/ is not resort-flagged (while the same type mis-filed is), and the length guidance splits by range for registries, by topic for everything else.
• Full suite 1282 passed; ruff clean; all 10 showcase lints green locally; copilot-instructions regenerated (drift guard green).

🤖 Generated with Claude Code

https://claude.ai/code/session_01X8L1A9vAa2WJKemtM21uR9

[Copilot]

Pull request overview

Introduces a new Registry page kind to ensure uniform enumerations (inventories, catalogs, rosters, code lists) are captured completely in the wiki with one cited row per entry, and adds a new graded corpus (werkhof) to validate this behavior end-to-end.

Changes:

• Add Registry as an OKF page type (rules + folder routing in okf.folder_for_type) and update curate guidance so oversized registries split by key range/subclass rather than topic.
• Add/extend rules docs (schema.md, core.md, genres/registry.md, and related briefs) to enforce per-entry citations, completeness, and promotion semantics.
• Add the new corpora/werkhof test corpus plus verify-corpus ground truth, and update CI/pages corpus floors to expect 10 corpora.

Reviewed changes

Copilot reviewed 24 out of 25 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
tests/test_okf.py	Adds routing tests for Registry → registries/ paths.
tests/test_curate.py	Adds tests ensuring registry pages aren’t resort-flagged and that length guidance is registry-specific.
README.md	Documents the new registries/ wiki folder.
corpora/werkhof/README.md	Documents the new corpus goals and expected registry/promotion outcomes.
corpora/werkhof/raw/wartungsbericht-2026-06.md	New prose source driving promotion, supersession, and open-points vs registry judgment.
corpora/werkhof/raw/stoercode-katalog.md	New enumerated fault-code catalog source (registry completeness + near-miss keys).
corpora/werkhof/raw/maschinenbestand-2026.md	New enumerated machine register source (registry completeness + per-row locators).
corpora/werkhof/raw/kundenliste.csv	New CSV roster source to grade “data rows are content” behavior.
corpora/werkhof/citadel.toml	Adds corpus workspace marker.
corpora/werkhof/.citadel_run.lock	Adds a runtime lockfile (should not be committed; blocks runs after checkout).
CLAUDE.md	Updates documentation to include the 10th corpus and werkhof in verify-corpus usage.
citadel/rules/tasks/reconcile.md	Extends “keep existing genre treatment” to include registry row lists.
citadel/rules/schema.md	Adds Registry to the type table and decision procedure; documents registry format contract.
citadel/rules/README.md	Indexes registries in the rules summary and adds the new genre brief.
citadel/rules/genres/registry.md	New genre brief defining completeness, per-row citations, promotion, and splitting guidance.
citadel/rules/genres/meeting-minutes.md	Routes perennial reference tables to registry treatment when uniformly enumerable.
citadel/rules/core.md	Clarifies that the granularity floor limits pages not rows; datasets aren’t “code structure.”
citadel/okf.py	Adds Registry → registries folder mapping.
citadel/curate.py	Makes page-length findings guidance type-aware for Registry.
CHANGELOG.md	Documents the Registry feature and its known limitation for oversized chunked sources.
.github/workflows/pages.yml	Bumps expected minimum corpus count from 9 → 10 for Pages build.
.github/workflows/ci.yml	Bumps expected minimum corpus count from 9 → 10 for showcase lint gating.
.github/copilot-instructions.md	Regenerates instructions to reflect the 10th corpus and updated verify-corpus usage.
.claude/skills/verify-corpus/werkhof/ground-truth.md	Adds the werkhof grading answer key.
.claude/skills/verify-corpus/SKILL.md	Updates verify-corpus skill docs to include werkhof and “ten corpora”.

[Copilot]

Pull request overview

Copilot reviewed 39 out of 40 changed files in this pull request and generated no new comments.

Comments suppressed due to low confidence (1)

citadel/okf.py:187

• The docstring references rules/genres/registry.md, but the rules files are generally referenced by their path within the rules tree (e.g. core.md, schema.md, genres/...). Using rules/... here is inconsistent and may be misleading for readers trying to locate the file.

    ``System``). ``Registry`` is a complete enumeration of like entries captured row-by-row from
    a source (a machine inventory, an error-code catalog, a customer roster — see
    ``rules/genres/registry.md``). ``Person``/``Organization``/``Project`` replace the old