Add native ClawDog round trip

[futureWA]

Summary

Adds native ClawDog report JSON-LD as a first-class xbrlkit projection and importer beside holon and TAVI, so authored ClawDog reports and converted filings can round-trip through XbrlModel and use the existing serve/view/tooling path.

Changes

• Adds to_clawdog / to_clawdog_report and from_clawdog_json / from_clawdog_report.
• Wires ClawDog into build, serve, export_filing, JSON sniffing, and view pass-through.
• Carries report/entity/period/unit/concept/fact/network structure, dimensions, calculation equations, fact provenance, concept references, parser fidelity fields (raw_value, source_hash), primary_document, and network role_id.
• Adds the LodgeiT equivalence fixture across ClawDog, TAVI, and holon, plus a whole-model ClawDog round-trip gate covering the review edge cases.

Output Impact

BROADER COVERAGE / CLI CONTRACT:

• xbrlkit build --format all now writes ClawDog output in addition to the existing projections.
• export_filing gains a clawdog format.
• serve / view can load and pass through .clawdog.jsonld reports.

Existing holon, TAVI, OIM, and property-graph output is intended to remain unchanged.

Testing

focused ClawDog gate -> 6 passed, 3 warnings
focused PR gate -> 53 passed, 99 warnings
full pytest -> 467 passed, 2 skipped, 162 warnings
ruff check -> All checks passed
ruff format --check -> 7 files already formatted
git diff --check -> clean

[futureWA]

xbrlkit PR 55 - native ClawDog round trip

[jfrench9]

Thanks Andrew — this is the PR I asked for and it is close. Before reading the asks, the measurements, because they are the good news.

What I checked, on this branch merged with current main (the branch is 11 commits behind; the merge is clean):

• Six holons from the corpus (3M 10-K 0000066740-25-000006, Apple 0000320193-23-000106, NVIDIA 0001045810-26-000021 and -000052, 0001193125-26-277521, and a RoboLedger-authored report) → to_clawdog → from_clawdog_json: periods, units, facts and networks come back identical on all six. Write is 0.8 s on the 3M filing, read 0.1 s.
• An Arelle parse of the 3M 10-K (3,150 facts, 559 networks, cubes included) → ClawDog → back keeps all 3,150 facts and 559 networks. The disclosures map and the segment information_block are byte-identical from the parse and from the ClawDog reload — so, unlike TAVI, this format carries the definition networks and every tool on main works over it.
• from_clawdog_json refuses a holon and a TAVI; json_kind classifies all three correctly.

So the shape is right. What follows is what stops the merge, then what I'd like, then nits.

Required

Three reader fallbacks overwrite a real empty value, and one field is dropped silently:

1. deserialize/clawdog.py:158 — reporting_style … or "clawdog". Every SEC filing comes back with reporting_style="clawdog"; the source has None. Keep None.
2. deserialize/clawdog.py:160 with serialize/clawdog.py:106 — the writer puts form or report_id in label, and the reader takes label as the form. A report with no form (the RoboLedger one) comes back with form == report_id. Write the form under lg:form only and stop reading label as a form.
3. deserialize/clawdog.py:253 — namespace … or (prefix if local else ""). _text("") is falsy, so a concept whose namespace is "" comes back with its prefix as the namespace: 426 of 1,329 concepts on the 3M filing (country:, cyd:, …). Only fall back when the key is absent, not when it is empty.
4. filing.primary_document is not written and not read, and it is not in the gap report either — the one thing the module docstring says will never happen. Write it, or name it in gaps.
5. README.md:210 now says the platform's SEC pipeline projects with to_clawdog_report. It does not — please revert that line. The serialize/deserialize READMEs are fine.
6. The PR body is the empty template. Output impact is not INTERNAL: build --format all now writes two more files and export_filing gains a format — that is additive coverage plus a CLI change, and it should say so. And name the filings you ran it against.

Suggested

• Add the whole-model gate. tests/test_deserialize.py::test_holon_round_trips_the_whole_model is the bar the holon meets; the ClawDog test compares a hand-picked signature on a three-fact fixture. Put the existing model fixture through to_clawdog / from_clawdog_json and compare field by field the way that test does. Items 1–4 above all fail it, which is the point.
• Empty the gap list. On a real parse the writer reports concept references, fact raw_value, fact source_hash and network role_id. Nothing in the format forbids carrying them — the format is defined in this module. Carry them and the corpus gate becomes model == back, the same fixed point the holon has. Related: serialize/clawdog.py:219 reports unmentioned concept for concept nodes the writer has just written unconditionally; that is not a gap. Drop it, or make it a note.
• In serialize/README.md the ClawDog row sits above holon. Put it last; the holon is the package's own projection.

Nits

• deserialize/clawdog.py:122 — the fallback sniff is "lg" in str(@context). Require documentType, or @context["lg"] == CLAWDOG_VOCAB.
• serialize/clawdog.py:33 — the xbrl prefix is declared and never used.
• The lg:AccountingEquation nodes restate every calculation arc, so the document is ~90% of the holon's size rather than "compact". Fine if you want them for ClawDog's own consumers; just not compact.
• Node ids are slugs under the vocab namespace (lg:fact:<slug>) while identity actually travels in lg:internalId. Works for our round trip (4,628 ids on 3M, no collisions), but another JSON-LD consumer gets slugs. If ClawDog has an identity scheme of its own, use it here.
• The doc-type IRI is under lodgeit.org with no document behind it. A sentence in the module docstring saying this module defines v1 would settle where the spec lives.
• periods.py:3–5 — "Four sources" and then "all three have to agree".

Ping me when it is pushed and I will merge and release it the same day.

[jfrench9]

Verified on the new head: whole-model round trip identical on the six corpus holons and on an Arelle parse of the 3M 10-K (references, raw_value, source_hash, role_id all carried; writer gaps empty), and the disclosures / information_block output is byte-identical from the parse and from the ClawDog reload. Thanks Andrew — merging.