#2412 Definitions with no typeVersion get the entire upgrade chain applied, which can corrupt current-shaped arguments

Direction: drop the legacy concept, three explicit states

None of the three options as filed is quite right, because they all accept the premise that an absent typeVersion is a meaningful signal worth preserving. Digging into the history says it is not.

Pre-CalVer was a bare integer, and the window was three days

typeVersion was introduced in #203 (df04d0dc, 2026-02-06) as z.number().int().positive().optional() — the values were 1, 2, 3. There was never a dotted pre-CalVer format. CalVer landed in #230 (9a418faa, 2026-02-09), after which modelDef.version is a CalVer string and nothing can write a number again. Four tagged builds fall in that window.

The restamp that swamp-club#900 just removed then ate the survivors: until c811f8ce, save() overwrote typeVersion on every write, so any numeric definition that was run, edited, or otherwise persisted in the last seven months was stamped to current CalVer. The surviving numeric population is definitions created in a three-day window in February that have not been saved once since.

Separately, the earliest upgrade entry in @swamp/issue-lifecycle — the only first-party chain — targets 2026.03.26.1, six weeks after CalVer. No genuinely legacy definition could need it.

Absence and malformation are different failures

The issue treats both as "no usable version". They are not. Absence means nobody recorded an intent — plausible, since definitions under models/ are git-tracked files people hand-write. A malformed value means someone recorded an intent and got it wrong, and silently ignoring it is the same class of defect as silently corrupting: we throw away something the author expected us to read.

The rule

Delete the "legacy pre-CalVer" concept entirely, including the numeric preprocess, and replace it with three states that each mean one thing:

This is option 1 for the case that actually occurs, with option 3's loud failure reserved for the case where a human demonstrably tried to say something.

Two supporting changes:

• The schema stops discarding numbers. A number coerces to its string form (1 → "1") rather than to undefined, so a malformed value survives into the object and model get can name it. Rejecting at the Zod boundary instead would make the definition unloadable, and model get is the command you reach for to diagnose exactly this file.
• resolveStaleness gains an invalid state alongside unknown. It still never throws — a read-only diagnostic has to survive the file it describes — but it reports the difference honestly, so model get says "malformed" rather than "legacy". unknown keeps its name and means unstamped.

swamp-club#900 shipped the DefinitionStaleness value object, the model get fields, and the method-run warning site a day ago, so this hangs on existing scaffolding rather than building new surface.

Cost

Fourteen sites carry the legacy framing, all comments, docs, and one test name, plus the preprocess in definition.ts. The only behavioural loss is that a surviving numeric definition stops auto-migrating — and it now fails loudly with a one-line fix rather than silently applying eighteen upgrades to arguments that may already be current-shaped.