docs(openusd): restructure run-on requirement prose

[marcschier]

What this changes

Readability repairs to workingdrafts/models/openusd/spec.md, produced in an
interactive Lingity editing session. Every change was an explicit editor
decision; nothing was applied autonomously.

Measured result

Policy architecture-review 1.3.0 (document HRI ≥ 85, every substantive block
≥ 85, zero high-severity findings), measured against the pre-edit bytes
(SHA-256 b9418c29…56524).

Measure	Before	After
Document HRI	3.30	4.21
Policy violations	61	32
Defects	449	419
High severity	49	28
Blocks below the 85 floor	11	3
Lowest substantive block	73.07	80.97

Severity mix: high 49 → 28, medium 223 → 212, low 177 → 179.

Principal repairs

• Abstract — split at the em-dash; in order to → to. The enabling verb
  allows is deliberately kept, so the registry is not newly asserted to
  perform storage, discovery and federation itself.
• §1 overview — the four-way discoverable / verifiable / versioned /
  federatable enumeration became a list, with the subject restated so it stays
  bound to each item. That block went 73.07 → 98.92.
• §5.4 asset resolution — procedure split out, with the
  validate-then-retrieve ordering kept explicit as "only after that validation".
• §5.1 — the xid definition split so the RFC 3986 constraints form their
  own sentence with xRegistry as the stated subject.
• §3 digest — the two obligations separated into one sentence each.
• Six further run-on sentences split in §1.2, §1.3, §4.4, §5.3, §5.5, §5.6.

Every split keeps the subject the source already stated, and preserves causal
and ordering cues with an equivalent connective rather than dropping them.

Protected-meaning result, stated honestly

The comparison against the original reports missing 22, added 29,
specified 1, disposition unresolved.

missing/added are re-segmentation artifacts. The claim extractor keys a
claim on a word-bag of its whole sentence, so splitting one sentence retires the
long claim and registers the shorter ones. Each pair was checked by hand and
carries the same obligation, actor and modality. Two are genuine clarity gains:
xRegistry also constrains… now states its subject, and they were derived
became the types were derived.

specified 1 is an authorized delta, not an accident:

actor specified: use assigned to 'consumer'

It comes from splitting "A Consumer … MUST verify it and MUST NOT use an
artifact whose digest does not match" into two sentences. The actor is the same
sentence's own subject — "A Consumer that has retrieved an artifact for which a
digest is declared" — so no obligation changed holder. This was put to the
editor as its own decision and approved with the noun repeated rather than a
pronoun.

Two candidate edits were rejected during the session for changing meaning:
naming an actor to satisfy the passive/agency rules, and rewriting the Abstract's
allows the storage… of into stores, manages, discovers and federates.

Residue — 28 findings retained by editor decision

Rule	Count	Reason
LING-NOMINALIZATION-001	9	Domain vocabulary: composition (USD composition arcs), dependency closure (a bolded defined term in §1.1), provenance, authentication, authorization, Section
LING-COMPOUND-DEPTH-001	6	Markdown link anchors read as hyphenated compounds; fixing breaks the table of contents
LING-PUNCTUATION-001	6	Markdown link parentheses and extension glosses such as (`.usda`, `.usdc`, `.usd`)
LING-ABBREVIATION-001	4	OPC/UA are the standards body's own names; MUST/SHALL/MAY/REQUIRED are mandated RFC 2119 boilerplate; USD is the spec's subject
LING-LIST-001	3	Line 200 is RFC 2119 boilerplate requiring byte-identical wording; line 489 is the normative assetkind enum value list

Remaining sub-floor blocks: 80.97, 81.69, 82.77.

This does not pass the policy

Final review: exit code 1, accepted: false, HRI 4.21 against a floor of 85.
The gap is a property of applying a plain-language policy to a normative
specification, not an editable defect. Editor acceptance of the residue is not a
Lingity pass, and this PR does not claim one.

Validation

• tools/verify.py . — succeeded, 51 files.
• pytest workingdrafts/models/openusd — 1 passed.
• git diff --check — clean; LF endings preserved, 0 CR bytes.
• Both commits signed off (DCO).

---

Update — actor naming (commit dde0bed)

A follow-up interactive session asked which actors this specification
specifies, how they are named, and whether better names exist.

What the audit found. No actor was defined anywhere. §2.1 declared no
capitalization convention and §2.2 defined only things. Eight roles were in
use, inconsistently:

Role	Spellings	Count	Bore a MUST
Consumer	Consumer / consumer	23 / 9	capitalized only
Producer	Producer	4	yes, §5.1.1 only
publisher	publisher	4	never
implementation	implementation	12	yes, 8 of them
client	client	1	MAY
resolver	3 surface forms	10	never directly

Consumer and Producer were capitalized like defined roles, while
implementation — carrying more obligations than Producer — was not.

Changes. Added §2.2.4 defining Consumer and Producer and stating the
capitalization convention; normalized the 9 lowercase consumer; renamed the 4
publisher to Producer; changed the single client to Consumer. The author
confirmed publisher and Producer name the same party — §5.1.1 had placed
Producer and implementation MUSTs five lines apart, which the text alone did
not resolve.

Three obligations gained an actor, each supported by surrounding text:

§	Before	After
4.1	"a collision MUST be rejected"	"An implementation MUST reject a case-insensitive collision"
4.5.5	"it MUST be published"	"a Producer MUST publish it"
5.1.1	"Inconsistent bindings MUST be rejected"	"A Producer MUST reject inconsistent bindings"

Measured:

	Before	After
LING-AGENCY-001	12	9
LING-PASSIVE-001	68	65
Actor-rule total	104	94
Defects	419	416
Mean block HRI	97.25	97.33
Blocks scoring 100	97	99

The composite document score moved 4.21 → 4.09 while every block measure
improved; that number is dominated by the 28 unchanged high-severity residue
findings and does not track this work.

protected_delta reports specified: 2 — reject assigned to 'implementation'
and reject assigned to 'producer'. Both were explicitly authorized, and both
actors come from surrounding text rather than invention.

Still open — not resolved. Three obligations keep no actor because the source
does not determine one: §1.3 "a usdassetid MUST NOT be derived from the
artifact's bytes" (by whom?), §1.3 "an xid … MUST NOT be used as an asset
identifier" (by whom?), and §5.1.1 "Metadata acquisition … MUST be explicitly
authorized" (by whom?). implementation also remains lowercase and undefined
while carrying 8 obligations, by author decision.

[marcschier]

Superseded — folded into xregistry#571 instead of shipping as a separate PR.

All three commits are now in that PR, rebased onto the fork's main (which is xregistry#571's head) as 3db72c4, b5c9abc, cf9f4a0. The rebased blob is byte-identical to what was reviewed here (70e63bce5ed6ca4582d0e1df816776b6156efa25), so nothing was lost in the move.

An upstream PR against xregistry/spec:main was attempted first and abandoned deliberately: upstream's copy of workingdrafts/models/openusd/spec.md is 854 lines against this branch's 964, because xregistry#571 still carries unmerged content for that file. Cherry-picking produced conflicts, and a three-way apply left five conflict regions, one spanning about 100 lines. Reconstructing the edits against upstream's older text would have produced a different change than the one measured, and would then have collided with xregistry#571 on the same file.

Closing without merging. No work is lost.