docs(data-integrity): domain integrity extends to OAS + codec types

[dimitri-yatsenko]

Expands the Domain integrity section of Data Integrity to cover what was only gestured at ("custom codecs attach domain rules … as well").

The added paragraph explains that Object-Augmented Schemas + the codec system carry domain integrity beyond scalar SQL types:

• A codec defines a domain-specific datatype (a movie, a sorting result, a fitted model) whose contents live in object storage but are addressed and governed by the schema.
• It can run arbitrary validation checks on encode (rejecting out-of-domain values, as enum does for a category) and expose typed access methods on decode.
• Object stores are not transactional participants: the codec validates and encodes before anything is written, the row commit is the atomic act that makes an object referenced, and a failed check aborts the insert with nothing written; unreferenced objects are removed by garbage collection.

Net: domain integrity holds uniformly from int32/enum to arbitrary structured and object-backed types, under one consistency guarantee. Links to the OAS explanation, the custom-codec how-to, and the codec API spec.

[MilagrosMarin]

Verified the substance: validate() runs at table.py:1369, before encode() at :1395, so "the check runs before anything is written" holds. Link targets all resolve and the #object-augmented-schemas anchor exists.

Two things:

1. The PR description still has the pre-30eb545 claim. It says "the object write is bound to the same database transaction as its row … row and object commit together or not at all" — which is what that commit was written to correct, and the file now correctly says object stores are not transactional participants. Worth updating the description so a squash-merge message doesn't carry the inaccurate version.

2. [garbage collection](referential-integrity.md) points at a page that doesn't cover it. referential-integrity.md has no mention of garbage collection (its two "orphan" hits are about orphaned records / dangling FKs — a different concept, which makes the mislabel easier to trip over). ../how-to/garbage-collection.md looks like the intended target. Worth noting §4 of this same file already links that page as "Referential Integrity" — text matching target — so it's just this new paragraph's label/target pairing.

[dimitri-yatsenko]

Both items addressed: the garbage-collection link now points at ../how-to/garbage-collection.md (the referential-integrity mislabel is gone), and the PR description's transactional claim is rewritten to match what 30eb545 corrected in the file — object stores are not transactional participants; the row commit is the atomic act — so a squash message carries the accurate version. One-line commit; re-review requested since the push resets the approval.

[MilagrosMarin]

Both verified. The link now points at ../how-to/garbage-collection.md — target resolves, and you correctly left §4's existing [Referential Integrity](referential-integrity.md) alone since that pairing was already text-matching-target. Description reads accurately now, and "a failed check aborts the insert with nothing written" is exactly right given validate() precedes encode().

One micro-nit, not worth a push on its own: "validates and encodes before anything is written" — encoding is the object write, so strictly it's validate → encode (object written) → row commit. The clauses after it make the model clear, so it reads fine in context.

Re-approving.