colregs 0.3.2 → 0.3.4

package/data/version.json

@@ -1,3 +1,3 @@
 {
-  "version": "0.3.2"
+  "version": "0.3.4"
 }

package/docs/adr/0010-text-withheld-jurisdictions.md

@@ -102,6 +102,8 @@ data one.
   international law where a national body deliberately has none. That is
   about deltas, not licences; it survives this ADR untouched and is the live
   blocker on CEVNI. Read the two together or the wrong one gets blamed.
+  *(ADR 0018 has since supplied that mechanism; the bar is now "land with
+  your tombstones", not "wait".)*
 - `schema/rules.schema.json` carries the conditional: `text` required unless
   `text_status` is `withheld`, in which case it is forbidden and
   `withheld_reason` is required. Only `withheld_reason` is refused on a

package/docs/adr/0013-corpus-files-with-editions.md

@@ -53,6 +53,25 @@ anyone has checked, and the `us/inland@2014` edition is recalled, not
 verified, and says so. That is the honest cost of this option and the reason
 to look at both.
 
+## Considered and declined (amendment, 2026-09-16)
+
+Four designs additive on option B, declined for now rather than on
+principle — each recorded in requirements.md §10 as a timed gate, closing
+event the first national text whose edition cannot be determined:
+
+- **Validity intervals on editions (`in_force_until`).** (GATE-7)
+  `superseded_by` plus `in_force` already infers validity for every
+  edition on file; a dedicated field would duplicate that inference.
+- **A per-source publication record separate from the edition, FRBR
+  manifestation.** (GATE-8) `source` lives on the corpus and one edition
+  has one source per jurisdiction today; nothing distinguishes yet.
+- **A content digest per corpus for verification.** (GATE-9)
+  `edition_status` already flags an unchecked corpus; a digest adds
+  tamper/drift detection nothing on file has needed.
+- **A skeleton per edition, instead of one shared skeleton.** (GATE-10)
+  Low risk today: no amendment has ever renumbered a rule (GATE-1's
+  verified history). The trigger is the first one that does.
+
 ## Consequences
 
 - Breaking file layout (REQ-PKG-4): `data/rules.json` no longer carries

package/docs/adr/0018-jurisdiction-delta-is-a-merge-patch.md

@@ -0,0 +1,102 @@
+# ADR 0018 — A jurisdiction delta is an RFC 7396 merge patch, and a tombstone is its `null`
+
+Date: 2026-09-16
+Status: proposed — merging this PR is the ruling; a revert undoes it
+
+## Context
+
+REQ-SCOPE-3 says a jurisdiction is a delta over `intl`: entries it does not
+mention are inherited. Q-11 found that unsafe as stated. 33 CFR 83 leaves
+Rule 28 "[Reserved]" and has no counterpart to 23(d)(ii), so a `us/inland`
+delta that was merely silent there would inherit `rule:28` and assert an
+international obligation the Inland Rules deliberately do not carry. Both
+absences are verified against the primary source (Cornell LII's copy of
+33 CFR 83: "§ 83.28 [Reserved] (Rule 28)"; 83.23(d) is a single paragraph
+with no subparagraphs). Q-11 asked for tombstones — "this entry deliberately
+does not exist here", distinguishable from "not yet transcribed" — and held
+every non-`intl` jurisdiction until the mechanism existed. Issue #45 asked,
+separately, which standard encoding the delta should follow.
+
+Both questions have been waiting on a second jurisdiction to force them.
+Issue #137 records that nothing was proposing one. `us/inland` already exists
+with two add-only entries under ADR 0008's carve-out, and the two absences
+above are the counterexample that carve-out was left open for: a delta that
+must suppress, not only add.
+
+## Decision
+
+1. **The delta semantics are RFC 7396, JSON Merge Patch.** A jurisdiction's
+   delta over `intl` is, in meaning, a merge-patch document keyed by entry
+   id: present keys are that jurisdiction's own entries, `null` keys are its
+   tombstones, and every key the patch does not mention is inherited. Any
+   conformant merge-patch implementation applied to the `intl` entries by
+   id reproduces the jurisdiction's resolved rule set. RFC 6902 is not
+   adopted: an operation script with JSON Pointer paths would duplicate the
+   id addressing `docs/identifiers.md` already provides.
+
+2. **Storage stays columnar; the patch is derived.** Entries keep their
+   `jurisdiction` field in the one `entries[]` table, as ADR 0008 laid them
+   down. Tombstones are a second top-level table, `suppressions[]`, one
+   record per (jurisdiction, entry): `jurisdiction`, `suppresses` (the
+   `intl` entry id), `cite` (the paragraph that has no counterpart) and
+   `why` (one sentence and the source). `test/data.test.mjs` builds the
+   merge patch from those two tables, applies it with a literal RFC 7396
+   implementation, and asserts the result equals the evaluator's
+   jurisdiction filter. That test is the statement that storage and
+   semantics agree; if they ever diverge, the test names which.
+
+3. **Replace is suppress plus add.** A jurisdiction that reads a paragraph
+   differently tombstones the `intl` entry and adds its own, under its own
+   id (ADR 0015 — the ids are distinct because the norms are). There is no
+   third operation and no in-place edit of an `intl` entry: REQ-SCOPE-4
+   holds.
+
+4. **The array caveat is accepted and irrelevant here.** Merge Patch
+   replaces an array wholesale. The patch is keyed at entry granularity, so
+   the only arrays it ever touches are inside a whole entry being added, and
+   an entry is small. A jurisdiction never patches one element of an
+   inherited entry's `lights`; it replaces the entry (point 3).
+
+5. **Only an inherited entry can be tombstoned.** A tombstone naming a
+   jurisdiction's own entry, or another jurisdiction's, is rejected by test:
+   the first is a deletion, the second was never in force.
+
+6. **A tombstone is verified the way an entry is.** Each needs one fixture
+   under its jurisdiction whose facts match the suppressed entry's predicate
+   and whose expectation omits it, and one `intl` fixture showing the entry
+   in force at the base (REQ-VERIFY-3, both sides).
+
+7. **Two tombstones land with this ADR**, both verified: `us/inland`
+   suppresses `rule:28` (Rule 28 "[Reserved]") and `rule:23d_ii` (no
+   counterpart to the under-7 m, 7 kn exception). They are enough to
+   exercise inherit and delete; no verified replace case exists in Part C
+   yet, so that branch of point 3 is stated, schema-ready and unexercised.
+
+## Consequences
+
+- **Q-11 is answered and its hold is lifted.** REQ-SCOPE-3's "no non-`intl`
+  jurisdiction lands before an explicit suppression mechanism exists" is
+  discharged by this mechanism; a jurisdiction lands with its tombstones or
+  it does not land. ADR 0008's add-only carve-out survives as the trivial
+  case: a delta with no tombstones is a merge patch with no `null`s.
+- **Issue #45 is closed by this ADR.** README names the RFC beside the
+  jurisdiction paragraph so a consumer knows which library semantics
+  reproduce the inheritance.
+- **A suppressed entry can leave a record out of vocabulary.** Under
+  `us/inland` a record carrying `activity:cbd` now selects nothing:
+  `rule:28` is gone and the Rule 23(a) entries it imported read
+  `activity:none`. That is the honest answer — "constrained by her draft" is
+  not an Inland status — and the package does not pick a reading. It is
+  recorded in `known_omissions` and pinned by a fixture. Whether the fact
+  vocabulary should carry per-jurisdiction membership so a consumer's
+  decode can be checked is a new question, not decided here.
+- **`rule:23d_i` cites a path Inland does not spell.** 33 CFR 83.23(d) has
+  no `(i)`; the content sits at bare `83.23(d)`. The entry is inherited
+  correctly — the norm is the same — and the citation-spelling question is
+  GATE-1's, untouched here.
+- **GATE-1, Q-10 and Q-8 are not taken.** Q-11 named a "second-jurisdiction
+  bundle"; this ADR takes only the item that gated data. The others stay
+  open on their own triggers.
+- **Cost to reverse.** Delete `suppressions[]` from data, schema and the
+  three tests; the two fixtures fail and are deleted with them. No consumer
+  reads the table yet. Pre-1.0, a revert.

package/docs/adr/0019-relation-reach-and-import-reads.md

@@ -0,0 +1,106 @@
+# ADR 0019 — What a relation reaches, and what an import reads
+
+Date: 2026-09-16
+Status: proposed — merging this PR is the ruling; a revert undoes it
+
+## Context
+
+REQ-MODEL-7 and the README define the six relation verbs by meaning and
+leave composition to the consumer (REQ-CONS-3). colregs-engine#33 checked
+the seven composition decisions its evaluator makes where the data is
+silent. ADR 0007 settled the largest, direction: a directed "this one
+prevails" is `rel:overrides`, and `rel:excludes` is pick-one between
+alternatives. Three readings of the verbs remain stated only in
+colregs-engine's `docs/engine-notes.md` (items 2, 3 and 5) and pinned only
+by its tests: how far a relation reaches, what a `one_of` yields, and which
+part of a referenced entry's predicate an import consults. Each is a
+reading of the Rules, and the package that carries the Rules should say it.
+
+## Decision
+
+1. **Displacing relations reach entries in force; composing relations
+   reach what is exhibited.** An entry is *in force* when its own predicate
+   holds. `rel:overrides` and `rel:exempts` act on entries in force and on
+   nothing else: an import is not in force, it is a light set its carrier
+   exhibits, and the carrier is what a displacing relation must name.
+   `rel:in_lieu_of`, `rel:excludes`, `rel:includes` and
+   `rel:conditional_includes` act on what one display may contain, imports
+   included, and remove nothing from the set of entries in force. Direction
+   is carried by the verb, never by modality: an override fires from a
+   forceful entry (ADR 0007), an exemption from a `modality:exempt` one,
+   and `rel:excludes` fires from nowhere — it is symmetric and already
+   barred to forceful entries by CI (`test/data.test.mjs`, "rel:excludes is
+   reciprocated"). A consumer that infers a veto from an excluder's modality
+   is reading a relation the data does not carry.
+
+2. **A `one_of` yields exactly one option per display, or none when the
+   carrier is `modality:may`.** Under a `may` carrier the chosen option is
+   exhibited in lieu of the carrier's own lights, and the none choice
+   exhibits them (25(d)(ii): the sailing lights, or failing them the torch).
+   An option already in force by its own predicate discharges the set: the
+   carrier imports nothing, and the option composes as the entry it is. A
+   vessel restricted in her ability to manoeuvre at anchor shows 30(a) or
+   30(b) because 30(a) and 30(b) apply to her, not because 27(b)(iv)
+   redirects her to them. Every choice is returned (REQ-MODEL-8).
+
+3. **An import reads the referenced entry's lights, their modality, and
+   its scalar gates — never its axes.** An import is a `rel:includes`, the
+   `rel:includes` of a `rel:conditional_includes` branch, or a `one_of`
+   option. `facts.json`'s `axes` and `modifiers` are the field: a key
+   declared there is the vessel's situation, and the carrier's redirect has
+   already placed her in it — 30(d) sends a vessel aground to "the lights
+   prescribed in paragraph (a) or (b)", and that she is not at anchor is the
+   premise of the redirect, not a gate on it. A key not declared there is a
+   fact the redirect does not alter, and the referenced entry's condition
+   on it binds: 30(b)'s "less than 50 metres" binds a vessel aground as it
+   binds one at anchor, and 25(b)'s "less than 20 metres" binds a vessel
+   under oars. Modality is the referenced entry's, resolved against the
+   vessel (Rule 28's three reds are `may`; the Rule 23 lights it imports
+   stay `shall`). A gate the carrier needs that the source does not carry is
+   written on the carrier's branch `when`, as 27(f) and 29(a) do; nothing
+   is inferred from a source's axes in either direction. REQ-MODEL-7's
+   "lights only, never its predicate" is amended to say this.
+
+4. **Two entries in force whose `rel:in_lieu_of` targets intersect are
+   alternatives to each other** and never share a display. This is an
+   invariant of the verb, not a declared pair list. The two pairs the data
+   can hold in force together read that way in the text: 23(d)(ii)'s
+   display is 23(d)(i)'s with the sidelights made practicable-only, and
+   25(d)(i)'s torch is what a vessel shows *if she does not* exhibit (a) or
+   (b). The Rule 24 entries that share targets — 24(a)(i) below and above
+   200 m, and 24(c) — are disjoint on their own predicates and never meet.
+   A future pair that lawfully combines is the evidence to reopen this
+   point; until one exists, the target sets are the declaration.
+
+Three shapes were considered and not adopted. Restating each option's
+scalar gate on the carrier's branch (30(d) split at 50 m) keeps "lights
+only" literal at the cost of transcribing 30(b)'s condition into 27(b)(iv),
+27(f), 29(a) and 30(d), and 25(b)'s into 25(d)(ii), plus a CI check to
+police the copies; the text redirects, it does not restate, and REQ-MODEL-5
+puts a gate where its fact lives. Declaring the point-4 pairs as a relation
+duplicates what the target sets say. A per-verb "reaches imports" flag
+would be set the same way on every entry, and a field nothing varies is
+decoration.
+
+## Consequences
+
+- **No data changes.** The README relation table, REQ-MODEL-7 and the
+  `relations` notes in `data/applicability.json` say the four points in
+  this PR. `docs/identifiers.md` and the schema are untouched.
+- **colregs-engine implements this in one issue, three deletions.** Its
+  modality-inferred exclusion path goes: `rel:excludes` is a co-occurrence
+  check on a display and nothing more. Its `includeApplies` goes: it
+  consulted one axis by name on `rel:includes` imports, the inverse of
+  point 3, and no import in the data reaches it on a coherent record —
+  every carrier either shares the source's position gate (23(b), 23(c),
+  25(c), 28), gates its branch on it (24, 27(f), 29(a)), or reads
+  `fact:making_way`, which refines `position:underway` (26(b)(iii),
+  26(c)(iii), 27(a)(iii), 27(b)(iii)). The availability test it already
+  applies to `one_of` options becomes the one rule for every import. Its
+  `docs/engine-notes.md` items 2–5 become pointers here. colregs-engine#33
+  items 1, 6 and 7 stay the engine's: presentation and output shape.
+- **A `one_of` option's scalar gate is a normative read of the source
+  entry.** Editing 30(b)'s length gate changes what a vessel aground may
+  show. That is the intended coupling: one paragraph, one gate.
+- **Cost to reverse.** Delete this file and restore two sentences; the
+  engine's deleted paths come back by revert. Pre-1.0, a revert.

package/docs/adr/0020-skeleton-is-a-delta.md

@@ -0,0 +1,80 @@
+# ADR 0020 — The skeleton is a delta too: a jurisdiction's paragraphs are a merge patch over `intl`
+
+Date: 2026-09-17
+Status: proposed — merging this PR is the ruling; a revert undoes it
+
+## Context
+
+ADR 0018 made the entries table a delta over `intl`. The skeleton was not
+one. `data/rules.json` keyed every paragraph by bare path, one record per
+key, and every one of them was `intl`; the corpus test asserted that a
+corpus paragraph's path belongs to the corpus's own jurisdiction. So a
+`us/inland` corpus could not carry `21(a)`: the key existed and was
+`intl`'s. The 2026-08-30 diff of 33 CFR 83 against Part C found 18 paths
+where Inland spells the same path with different text, three it does not
+spell at all, and a dozen it adds. None of the 18 could be stated. That,
+not a licence, is why the Inland corpus stub had no paragraphs (issue
+#161), and why every planning session stalled on "the second jurisdiction".
+
+Solace ruled on issue #137 on 2026-09-17: `us/inland`, Part C, before v1.
+
+## Decision
+
+1. **`rules.json` gains `deltas`, keyed by jurisdiction.** A delta has
+   `paragraphs` (own rows, keyed by path, each carrying its jurisdiction)
+   and `suppressions` (paths the source does not spell, each with a `why`).
+   `paragraphs` at the top level stays the `intl` base, untouched, so every
+   existing consumer reads what it read.
+
+2. **A jurisdiction's resolved skeleton is the RFC 7396 merge patch of its
+   delta over the base**, the same statement ADR 0018 makes for entries: own
+   rows present, suppressed paths `null`, an unmentioned path inherited
+   whole. A restated path merges field by field, so an own row that says
+   nothing about `images` inherits the base figure. The test builds the
+   patch, applies a literal RFC 7396 implementation, and asserts the
+   resolved skeleton is what corpora, entries and fixtures key into.
+
+3. **A path is stated in a delta when its text differs from `intl` or has no
+   `intl` counterpart; suppressed when the source does not spell it; silent
+   otherwise.** A `[Reserved]` section is a spelled path with reserved text,
+   so it is stated, not suppressed: Inland `28` and `26(d)` are rows whose
+   corpus text will be `[Reserved]`.
+
+4. **An entry in force under a jurisdiction cites a path in that
+   jurisdiction's resolved skeleton.** An inherited `intl` entry whose path
+   the delta suppresses cannot stand by silence: it is tombstoned, and where
+   the norm survives under another path, replaced (ADR 0018 point 3).
+
+5. **`us/inland` lands with its Part C skeleton delta**, every row from
+   `docs/verification/2026-08-30-q6-q8.md`: 19 restated paths (the 18
+   differing plus `28`), 18 Inland-only paths, and `23(d)(i)`, `23(d)(ii)`,
+   `23(d)(iii)` suppressed. A new sub-paragraph is stated when `intl`
+   already carries sibling sub-rows of that parent; where `intl`'s parent is
+   one row, the parent is restated and the sub-paragraph waits for the
+   corpus. So `24(f)`'s Inland `(iii)` and `22(d)`'s Inland `(ii)` wait
+   (`24(f)`, `22(d)` are one row each in `intl`), while `24(g)(v)` is stated
+   (`intl`'s `24(g)(i)` is a sibling row). `NRHB_23_e.png` and
+   `NRHB_24_j.png` are mapped to `23(e)` and `24(j)` and stop being
+   unmapped.
+
+6. **The first replace lands with it.** Inland 83.23(d) is the under-12 m
+   provision as one paragraph; `23(d)(i)` is not a path there. Point 4
+   forces the branch ADR 0018 left unexercised: `rule:23d_i` is tombstoned
+   under `us/inland` and `rule:23d`, same predicate and lights, cites
+   `23(d)`. Fixtures show both sides.
+
+## Consequences
+
+- The Inland corpus can now be filled: every path it will carry resolves.
+  Text is 17 U.S.C. 105, so REQ-PROV-2 does not gate it. That is the next PR.
+- `us/inland` is now six applicability records and a skeleton delta of 40
+  rows. README's coverage statement says so; it is still not a model of the
+  Inland Rules until the entries for the restated paths land.
+- The skeleton edition `us/inland@2014` is declared and still recalled, not
+  verified against the Federal Register (REQ-LANG-10). The corpus fetch
+  verifies it or corrects it.
+- GATE-1 is untouched: a path that differs across jurisdictions at one time
+  is REQ-MODEL-1 working as designed, and this ADR gives it a home.
+- **Cost to reverse.** Delete `deltas` from data and schema, the resolver and
+  three tests, `rule:23d` and its tombstone, two fixtures; unmap two images.
+  No consumer reads `deltas` yet. Pre-1.0, a revert.

package/docs/budgets.json

@@ -23,9 +23,12 @@
     "docs/adr/0015-rule-ids-are-paragraph-keys.md": 210,
     "docs/adr/0016-encounter-roles-are-pooled-across-frames.md": 90,
     "docs/adr/0017-closed-vocabularies-are-prefixed-identifiers.md": 120,
+    "docs/adr/0018-jurisdiction-delta-is-a-merge-patch.md": 120,
+    "docs/adr/0019-relation-reach-and-import-reads.md": 110,
     "docs/timeline.md": 130,
     "docs/maritime-sources.md": 130,
-    "docs/decisions.md": 40
+    "docs/decisions.md": 40,
+    "docs/adr/0020-skeleton-is-a-delta.md": 110
   },
   "json_prose": {
     "targets": [

package/docs/decisions.md

@@ -3,4 +3,8 @@
 Rulings from `/sequence` sessions and other closed judgment calls. One line
 each: date, short name, the answer, a link to the argument.
 
+- 2026-09-16 — docs ownership: colregs-engine's docs are sited wrong; colregs owns docs going forward, no new doc lands in colregs-engine. `normative-language.md` was already migrated the same day (commit fe0f8c2, #125); migration of the remaining three (`formal-methods-glossary.md`, `formal-methods-reading-list.md`, `engine-notes.md`) is not yet scheduled. Argued on kanban; no public writeup to link.
 - 2026-09-16 — own vs self: rename the subject segment `own` to `self` across the situation record (fact keys, precedence effect keys, facts.json, fixtures, docs/identifiers.md). Work spawned as [colregs#138](https://github.com/mark-brannan/colregs/issues/138).
+- 2026-09-16 — Q-54, Rule 17's phases are monotone: once 17(a)(ii) permission or the 17(b) duty has arisen for the stand-on vessel, it does not fall back on belated give-way compliance. Argued in [colregs#72](https://github.com/mark-brannan/colregs/issues/72); recorded in `docs/requirements.md`.
+- 2026-09-16 — Q-53, 17(a)(ii) is an exception, not a suspension: the stand-on vessel's duty stands and a departure is lawful only as action to avoid collision; a monitor flags any other alteration. Argued in [colregs#72](https://github.com/mark-brannan/colregs/issues/72); recorded in `docs/requirements.md`.
+- 2026-09-17 — second jurisdiction: `us/inland`, Part C, before v1. Ruled on [colregs#137](https://github.com/mark-brannan/colregs/issues/137); skeleton delta lands under ADR 0020, text and entries follow.

package/docs/gates.json

@@ -11,6 +11,7 @@
     "first-non-english-corpus": "the second corpus of any one jurisdiction, which in practice is the first non-English text",
     "first-community-translation-of-national-corpus": "the first `community`-tier translation of a `national`-tier corpus",
     "first-external-contribution": "the first merged contribution from someone other than the copyright holder",
+    "first-undated-national-edition": "the first national text whose edition cannot be determined",
     "none": "the decline does not get cheaper or dearer with time; recorded so it is not re-read as merely deferred"
   },
   "statuses": {
@@ -79,6 +80,46 @@
       "status": "open",
       "settled_by": null,
       "note": "ADR 0004 settles the code licence (MIT \u2192 Apache-2.0) and leaves the data licence open. Already-published npm versions stay under the licence they shipped with; the gate governs future releases only, and is held open deliberately by REQ-PROV-7's CONTRIBUTING.md terms (DCO-style certification plus a relicensing grant); a CLA-assistant bot is the upgrade path if contributors arrive."
+    },
+    {
+      "id": "GATE-7",
+      "title": "validity intervals on editions (`in_force_until`)",
+      "declined_in": "docs/adr/0013-corpus-files-with-editions.md",
+      "closing_event": "first-undated-national-edition",
+      "trigger": "an edition superseded with no successor edition registered to bound it \u2014 repeal without replacement, or two editions of one jurisdiction whose validity ranges cannot be ordered from `superseded_by` alone",
+      "status": "open",
+      "settled_by": null,
+      "note": "Declined alongside ADR 0013 (option B, GATE-2 adopted): `superseded_by` plus `in_force` infers validity for every edition on file today. An explicit `in_force_until` is additive, not a rewrite, so paying for it before a case needs it buys nothing."
+    },
+    {
+      "id": "GATE-8",
+      "title": "a per-source publication record separate from the edition (FRBR manifestation)",
+      "declined_in": "docs/adr/0013-corpus-files-with-editions.md",
+      "closing_event": "first-undated-national-edition",
+      "trigger": "two distinct published sources for the same edition that disagree \u2014 e.g. an official gazette printing and a codifier's republication of the same edition differing in errata \u2014 so \"the source\" stops being a property of the edition",
+      "status": "open",
+      "settled_by": null,
+      "note": "Declined alongside ADR 0013: today `source` lives on the corpus and one edition has one source per jurisdiction. A manifestation layer is additive when a second publication record is actually needed."
+    },
+    {
+      "id": "GATE-9",
+      "title": "a content digest per corpus for verification",
+      "declined_in": "docs/adr/0013-corpus-files-with-editions.md",
+      "closing_event": "first-undated-national-edition",
+      "trigger": "a corpus found to differ from its cited source with no accompanying edition or source change \u2014 silent drift a hash would have caught and prose review did not",
+      "status": "open",
+      "settled_by": null,
+      "note": "Declined alongside ADR 0013: `edition_status` (verified/claimed/unknown, this ADR's before-merge item 1) already flags an unchecked corpus. A digest adds tamper/drift detection on top, which nothing on file today has needed."
+    },
+    {
+      "id": "GATE-10",
+      "title": "a skeleton per edition",
+      "declined_in": "docs/adr/0013-corpus-files-with-editions.md",
+      "closing_event": "first-undated-national-edition",
+      "trigger": "the first amendment that deletes or renumbers a path",
+      "status": "open",
+      "settled_by": null,
+      "note": "Declined alongside ADR 0013: one shared `data/rules.json` skeleton is low risk because no amendment has ever renumbered a rule (seven IMO amendments since 1972, none doing so). A per-edition skeleton is the fallback the day one does."
     }
   ]
 }

package/docs/identifiers.md

@@ -181,7 +181,7 @@ is an enumerated fact, so its values follow the bare-fact-name rule above:
 Rule 13(d) is the reason history is a class and not a note. Once a vessel is
 overtaking, a subsequent alteration of the bearing does not make her a
 crossing vessel; the instantaneous geometry, read alone, says otherwise and
-hands the duty to the wrong vessel. So the latch is a fact:
+hands the duty to the wrong vessel. So overtaking history is a fact:
 
 - `self:hist:was_overtaking` — this subject was, earlier in this encounter,
   an overtaking vessel with respect to the other.
@@ -433,6 +433,9 @@ implemented in the reference evaluator and asserted by the fixtures.
   words. It is deliberately not a light reference and does not resolve to
   one.
 
+Not being an identifier says nothing about whether a vocabulary gets a
+catalog label; see REQ-LANG-6.
+
 ## This does not reverse GATE-5
 
 GATE-5 declined a CI-enforced terminology glossary, permanently, for the

package/docs/maritime-sources.md

@@ -26,6 +26,12 @@ question the rule text leaves open.
 - **Nautical Institute — *Action by the Stand-On Vessel*** (Seaways case study). The stand-on vessel's stages as taught; does not reach whether they are reversible.
   <https://www.nautinst.org/resources-page/200115-action-by-the-stand-on-vessel.html>
 - **USCG Navigation Rules (Amalgamated)**. <https://www.navcen.uscg.gov/navigation-rules-amalgamated>
+- **Steamship Mutual — *Navigation in Restricted Visibility***. Confirms Section II reapplies once vessels are in sight of one another again; silent on whether status accumulated before the loss of sight (a `Q-51`/`Q-52` classification, a `Q-54` Rule 17 stage) carries forward.
+  <https://www.steamshipmutual.com/publications/articles/restrictedvisibility0109>
+- **Britannia P&I — *Navigation in restricted visibility***. No stand-on vessel exists under Rule 19; also silent on status continuity across the transition.
+  <https://britanniapandi.com/en_gb/guidance/navigation-in-restricted-visibility/>
+- **Professional Mariner — *Overtaking or crossing? Don't assume what other ship will do***. The Rule 13(d) bearing-shift overtaking history, discussed only within continuous visual contact.
+  <https://professionalmariner.com/overtaking-or-crossing-dont-assume-what-other-ship-will-do/>
 
 ## Marine incident analysis
 
@@ -47,6 +53,16 @@ ambiguous give-way/stand-on roles are what the Rules are written against.
   <https://www.bahamasmaritime.com/wp-content/uploads/2026/02/2026-5-Polesie-Verity-ReportAndAnnexes.pdf>
 - IMO GISIS Marine Casualties and Incidents module. Not a paper but a source class: the mandatory-reporting database of marine safety investigation reports. Ground truth for real COLREGS-relevant incidents.
   <https://www.imo.org/en/OurWork/IIIS/Pages/Marine-Safety-Investigation-reports.aspx>
+- NTSB/MAR-16/01 — *Conti Peridot*/*Carla Maersk* collision, Houston Ship Channel. Occurred entirely within restricted visibility; no in-sight phase, no Rule 13/17 discussion. Checked for `Q-55` and found not on point.
+  <https://www.ntsb.gov/investigations/AccidentReports/Reports/MAR1601.pdf>
+- Dutch Safety Board — *Collision in Restricted Visibility* (*VB Seal*/*Gulholmen*, Maasmond, 2026). Rules appendix cites only Rules 6 and 19; the encounter never had an in-sight phase. Checked for `Q-55` and found not on point.
+  <https://onderzoeksraad.nl/wp-content/uploads/2026/07/collision-in-restricted-visibility.pdf>
+- TSB Canada M23C0143 — ferry *Svanoy*. Fog/lookout casualty; no Section II status question arises.
+  <https://www.tsb.gc.ca/eng/rapports-reports/marine/2023/m23c0143/m23c0143.html>
+- SWZ Maritime on Nautical Institute MARS 202601 (near-miss in fog). Operational-failure summary only; no rule-regime discussion.
+  <https://swzmaritime.nl/news/2026/02/13/near-vessel-collision-in-fog/> · MARS index: <https://www.nautinst.org/technical-resources/mars.html>
+- Marine Insight — fog/TSS-turn collision write-up. Overtaking occurred entirely within fog; no visibility transition at issue.
+  <https://www.marineinsight.com/real-life-incident-vessels-collide-in-dense-fog-during-sudden-tss-turn/>
 
 ## Known gaps
 
@@ -55,4 +71,5 @@ ambiguous give-way/stand-on roles are what the Rules are written against.
   manoeuvrability, shoal water, a lee shore, set, visibility or sea state.
 - Two standard texts — Cockcroft & Lameijer, *A Guide to the Collision Avoidance Rules*, and Farwell's *Rules of the Nautical Road* — are not online. Either may settle how the stages of a close-quarters encounter are divided, and whether they are treated as irreversible.
 - Several of the questions in `docs/requirements.md` §11 appear unlitigated: overtaking geometry with no risk of collision, an overtaking situation becoming a head-on, resumption of course by a stand-on vessel that has acted, and the fate of accumulated Section II state across a visibility transition.
+- `Q-55` (visibility transition): checked casualty reports, P&I guidance and trade-press commentary beyond the sources above — no authority found in either direction on whether a `Q-51`/`Q-52` classification or a `Q-54` Rule 17 stage survives a restricted-visibility interruption or resets on regaining sight. Deferred indefinitely, 2026-09-16; Cockcroft & Lameijer and Farwell (both already listed above as unreachable) remain the most plausible sources to settle it.
 - BAILII refuses automated access. Use the National Archives Find Case Law service: <https://caselaw.nationalarchives.gov.uk/>

package/docs/part-b-invariants.md

@@ -67,7 +67,7 @@ is Rules 11–18, III is Rule 19; which governs a pair is `INV-19a-scope`.
 # Rule 13 — Overtaking
 
 13(a) assigns a role, 13(b) deems the encounter, 13(c) resolves doubt, 13(d)
-latches the result against the geometry that follows.
+persists the result against the geometry that follows.
 
 ### INV-13b — the overtaking deeming test
 
@@ -122,7 +122,7 @@ well as Rules 14 and 15.
   Settled by P4.2's role assignment: the override is or is not needed to keep
   "never both give-way" true.
 
-### INV-13d — the latch: the classification survives the geometry
+### INV-13d — overtaking history: the classification survives the geometry
 
 `13(d)` · classification · temporal, a state to every later state of the
 encounter · `✎`
@@ -140,10 +140,11 @@ until self is finally past and clear, the encounter type of the pair is
   forbids: an overtaking drawn out on the bow reads *crossing*, and hands
   give-way to the wrong vessel.
 - **Undetermined term.** "Finally past and clear" has no fact and no threshold;
-  the data's latch is consumer-cleared and never self-clears (`Q-47`), a trace
-  model's is its segmentation, and no bearing or range is invented here.
-- **Readings in doubt.** Two: what arms the latch, `Q-51`; what it forbids,
-  `Q-52`.
+  the data's overtaking history is consumer-cleared and never self-clears
+  (`Q-47`), a trace model's is its segmentation, and no bearing or range is
+  invented here.
+- **Readings in doubt.** Two: what arms the overtaking history, `Q-51`; what it
+  forbids, `Q-52`.
 - Entries: `rule:13d`, reading `self`/`other:hist:was_overtaking` and no geometry; the
   `was_overtaking: false` gates on `rule:14b`, `rule:15a:crossing`, `rule:15a:keep_out_of_the_way`
   implement `Q-52`'s *broad* reading. Settled by `Q-51`, `Q-52`, then TLC on a
@@ -162,8 +163,8 @@ type alone cannot distinguish `Q-52`'s readings.
 
 - **State remembered.** As `INV-13d`.
 - Entries: `rule:13a`'s `any_of` second limb, `self:hist:was_overtaking: true` — role
-  asserted from the latch directly, keeping the limbs separable. Settled by
-  `Q-52`.
+  asserted from the overtaking history directly, keeping the limbs separable.
+  Settled by `Q-52`.
 
 ---
 
@@ -174,10 +175,10 @@ type alone cannot distinguish `Q-52`'s readings.
 `14(b)` reading 14(a)'s conditions · classification · single-state · `✎`
 
 **Invariant.** At any state at which two power-driven vessels are in sight,
-both underway, with risk of collision, and neither latched under `INV-13d`, a
-head-on situation is deemed to exist exactly when each sees the other ahead or
-nearly ahead: `self:geo:rel_bearing_deg` and `other:geo:rel_bearing_deg` each
-in [0°, 11.25°] ∪ [348.75°, 360°).
+both underway, with risk of collision, and neither carrying overtaking history
+under `INV-13d`, a head-on situation is deemed to exist exactly when each sees
+the other ahead or nearly ahead: `self:geo:rel_bearing_deg` and
+`other:geo:rel_bearing_deg` each in [0°, 11.25°] ∪ [348.75°, 360°).
 
 - **Both bearings, not one.** 14(b)'s lights limb is a statement about the
   *other* vessel's aspect; one bearing alone swallows part of the crossing
@@ -219,9 +220,9 @@ way; both are directed to act, and the duty is symmetric.
 `15(a)` · classification · single-state · `✎`
 
 **Invariant.** At any state at which two power-driven vessels are in sight,
-both underway, with risk of collision, and neither latched under `INV-13d`, the
-encounter type is `crossing` exactly when it is neither `overtaking` under
-`INV-13b` nor `head-on` under `INV-14b`.
+both underway, with risk of collision, and neither carrying overtaking history
+under `INV-13d`, the encounter type is `crossing` exactly when it is neither
+`overtaking` under `INV-13b` nor `head-on` under `INV-14b`.
 
 - **Derived, not enumerated.** Enumerating the sector gives two statements of
   one boundary, and the partition's only failure mode is editing one;
@@ -395,7 +396,9 @@ without further action by the give-way vessel.
 - **Formally.** Reachability over the stand-on vessel's manoeuvring envelope:
   `Safe(s, σ_own, σ_other)` of ADR 0005 §5 with the give-way strategy set
   unconstrained — the first Part B invariant needing `kin:dynamics`.
-- **Reading in doubt.** Suspension or exception: `Q-53`.
+- **Reading ruled.** Exception, not suspension (`Q-53`, 2026-09-16): the
+  17(a)(i) duty stands and the only lawful departure is action to avoid
+  collision; a checker flags any other alteration.
 
 ### INV-17b — phase 3: the stand-on vessel shall act
 
@@ -462,11 +465,12 @@ is at each state in exactly one of three phases:
 
 Phase 3 dominates phase 2: an obligation is not qualified by a permission.
 
-- **Reading in doubt.** Monotone or not: `Q-54`.
-- **State remembered.** If monotone, one three-valued latch per stand-on vessel
-  per encounter; if not, no latch but `INV-17c` still needs the phase of each
-  action. TLC distinguishes them: under monotone a late-complying give-way
-  vessel still leaves stand-on action permitted; under reversible it does not.
+- **Reading ruled.** Monotone (`Q-54`): once phase 2 or 3 is reached for the
+  stand-on vessel in an encounter, it does not fall back on belated give-way
+  compliance.
+- **State remembered.** One three-valued marker per stand-on vessel per
+  encounter, monotone: a late-complying give-way vessel still leaves stand-on
+  action permitted.
 
 ---
 
@@ -601,7 +605,7 @@ not jointly exhaustive (`INV-19a-third-state`); Rule 19 supplements Section I
 
 - **Temporal note.** The switch is a function of the current state; a pair may
   cross it and back inside one encounter, and no paragraph says what becomes of
-  a 13(d) latch or a Rule 17 phase: `Q-55`.
+  a 13(d) overtaking history or a Rule 17 phase: `Q-55`.
 - **Undetermined term.** "In or near an area of restricted visibility" has no
   fact (3(l)'s atmospheric condition is nowhere in `data/facts.json`), so `rule:19a`
   selects Section III for any pair not in sight — wider — and records a gap.