chore(metamodel): make regeneration reproducible, and pin down the dropped-typeof bug

[kervel]

src/metamodel (linkml_meta) is generated by gen-rust from a linkml checkout, but nothing
recorded which generator revision produced the committed crate. That makes a regen
unreproducible and, worse, silently destructive: a regen from the wrong revision reverts whatever
the committed crate has that the generator no longer emits.

Not hypothetical — this was found while trying to regenerate. The per-key map merge strategy
(option_map_overwrite, which keeps a slot_usage override of annotations from dropping the
inherited ones, added in f79d7fe) was absent from the generator branch that was checked out; a
regen would have replaced it with whole-map overwrite_except_none and no test would have failed.

What this changes

regen.sh:

• resolves the repo root from the script location, and the sibling linkml checkout from the main
  working tree, so it works from a git worktree instead of silently looking for ../linkml next to
  the worktree;
• takes LINKML_DIR as an override;
• stamps src/metamodel/GENERATED_FROM with the generator revision used and with every
  generator commit not yet in linkml/main — while that list is non-empty, the crate cannot be
  reproduced from upstream linkml alone;
• grows --check, which regenerates into a temp dir, formats it the same way, diffs against the
  committed crate, and writes nothing.

README.md documents both, plus the rule they enforce: a generator fix goes upstream as a linkml PR
before the regenerated crate is committed here.

The bug this was gating, recorded as tests

LinkML spells the keys typeof and abstract. Both are Rust keywords, so the generated fields are
typeof_/abstract_, and nothing maps between the two spellings — serde drops the value without
error and a derived type inherits nothing:

type_ancestors(mainTrackLength)  ->  ["mainTrackLength"]
                          expected  ["mainTrackLength", "trackLength", "integer"]

determine_rdf_type_info is not at fault; its ancestor walk is correct and is being handed a
truncated list. Consequences: rdf_datatype_iri is not inherited, is_range_iri is not inherited
(so a type declared typeof: uri serializes as a literal instead of a named node),
is_integer/is_floating_point are false, and abstract: is dropped on four structs.

The fix is a generator change (linkml/linkml#3919) plus a regen, neither of which belongs in this
PR. So the behaviour is committed as seven tests that fail today and are #[ignore]d with a pointer
to #108 — five in schemaview/tests/typeof_inheritance.rs and two in
runtime/tests/turtle_custom_types.rs. Confirmed failing for the right reason:

geotest:mirror "https://mirror.brussels.be" ;   # should be <https://mirror.brussels.be>
geotest:track_length "1200" ;                   # should carry ^^xsd:integer

Un-ignore them in the commit that lands the regenerated metamodel.

Current --check output

Drift against the local generator stack, all of it the output of two open linkml PRs — the five
serde renames (#3919) and the *OrSubtype base-class variants (#3662). Everything else in the crate
reproduces byte-for-byte after formatting, Cargo.toml included.

Tracking, one issue per upstream PR so they close as they merge: #108 (#3919), #110 (#3662), #111
(#3661, no effect on this crate's output), with #109 for the regeneration itself.

Notes

• --check needs a local linkml checkout, so it cannot run in CI as-is.
• src/metamodel/Cargo.toml~ is a committed editor backup; --check excludes it rather than
  deleting it here.

cargo fmt --all --check and cargo test --workspace (72 suites) are clean.