Execution Pipeline

create and update run through a fixed sequence of phases. Knowing the order matters for two things in particular: what ctx.values() can see from inside a resolver (everything resolved in an earlier phase, nothing from a later one), and when validation fails "fast" -- most phases that can produce an error return immediately if one occurred, without running any later phase against already-invalid data.

The macro only generates the phases a given schema actually needs -- a schema with no #[constant] fields has no constant-attachment step at all, a schema with no #[depends_on] fields has no dependent-resolution loop, and so on. Nothing described below is dead code or a runtime branch for schemas that don't use a given feature.

create

1. Ignore -- evaluate #[ignore]/#[ignore_init] (field-level and grouped together), then apply #[ignore_init] overrides.
2. Required -- evaluate conditional #[required(...)] and check bare #[required] fields for a missing value. Fails fast.
3. Validate -- #[validate] runs for required/lax and virtual fields together, as one phase. Defaults for unset #[lax] fields are applied here too. Fails fast.
4. Re-validate -- #[re_validate], same grouping as validate, only for fields that validated successfully. Fails fast.
5. Post-validate (pre) -- every #[post_validate(...)] group's pre_validate handler, against a snapshot from before this phase. Fails fast.
6. Post-validate (main) -- every group's main validate handler; skipped entirely if step 5 produced any error. Fails fast.
7. Sanitize -- #[sanitize] on virtual fields that were provided and not ignored, once step 6 has succeeded.
8. Resolve dependents -- #[resolve], one round per dependency-graph level, looping until nothing new changes. ctx.values() inside a resolver reflects everything resolved in earlier rounds.
9. Attach constants -- #[constant], after dependents resolve, so a constant resolver can read resolved dependent values via ctx.values().
10. Attach timestamps -- #[created_at]/#[updated_at], after constants. The shared resolver is called at most once.
11. Prepare on_success/on_failure triggers for the returned handle.

update

1. Ignore -- evaluate #[ignore]/#[ignore_update], then apply bare #[ignore_update] overrides.
2. Nothing-to-update checkpoint 1 -- if no required/lax/virtual field actually submitted survives ignore/#[readonly] filtering, fail immediately with "nothing to update", before the required check even runs.
3. Required -- conditional #[required(...)] only (bare #[required] is creation-only). Fails fast.
4. Validate -- same as create's step 3. Fails fast.
5. Re-validate -- same as create's step 4. Fails fast.
6. Post-validate -- pre_validate then main validate, gated the same way as create's steps 5-6. Fails fast.
7. Evaluate update validity -- recompute the change set and drop any field whose value turned out unchanged, once, right after post-validate (raw_input() still shows what was actually submitted; input() reflects this filtering). Fails fast.
8. Nothing-to-update checkpoint 2 -- if nothing is left to change after step 7 and no virtual field is still relevant (its dependent(s) haven't resolved yet), fail immediately, before dependent resolution runs.
9. Sanitize -- same condition as create's step 7.
10. Resolve dependents -- one pass per dependency-graph level.
11. Nothing-to-update checkpoint 3 -- if the change set is still empty after dependent resolution, fail with "nothing to update".
12. Attach timestamps -- #[updated_at]/#[optional_updated_at].
13. Prepare on_success/on_failure triggers for the returned handle.

Why three "nothing to update" checkpoints?

Each catches a different way an update can turn out to be a no-op: submitting only fields that get filtered out before validation even starts (checkpoint 1), submitting a field with its current, unchanged value (checkpoint 2), or submitting a virtual field whose dependent resolves back to the value it already had (checkpoint 3, since a virtual field's relevance can't be known until its dependent actually resolves). All three surface the same way: update returns an error whose errors field is None -- not a validation failure, just nothing to do.

let err = DataInputModel.update(existing, updates, ()).unwrap_err();
assert!(err.errors.is_none()); // "nothing to update", not a validation error