Add a test suite to detect backwards-incompatibility of new changes - #28

write_reference writes a small RNTuple with the canonical value recipe
from reference_values.hpp. The recipe compiles unmodified against the
headers of every released tag: headers added later are guarded with
__has_include and members added later with member detection inside
function templates, so the same source can backfill reference files for
old tags and write the current one at release time. Values match the
SHiP::test::make* generators for members that existed at v0.1.0, with
distinctive values for later members so they are exercised too.

Assisted-by: claude-code:claude-fable-5

One frozen reference RNTuple per released version, written by the tag's
own headers (which define the on-disk schema) via the version-agnostic
writer, plus reference_head.root tracking main (regenerated with
'pixi run update-reference-head' whenever the event model changes). The
backfilled files were all written by the ROOT pinned in pixi.lock
(6.40.02), so they freeze the historical schemas but not the historical
writer-ROOT versions; files for future releases are written at release
time and capture both. Frozen files are never rewritten.

Assisted-by: claude-code:claude-fable-5

One ctest test per committed reference file, for clear failure
attribution. The reader opens the RNTuple without an imposed model, so
the on-disk descriptor and the current dictionary drive automatic
schema evolution: members map by name and members absent on disk come
back default-initialized. Expected values are the canonical recipe,
masked to the writing version (geometryNodeId for pre-v0.4.0 files,
RecParticle::hits for pre-v0.3.0); top-level field presence is also
checked against the version. A change that breaks reading of existing
data now fails CI.

Assisted-by: claude-code:claude-fable-5

schema_snapshot dumps the persistent schema of every dictionary class
(TClass version, checksum and members, plus the RNTuple field tree of a
probe model) and compares it against the committed snapshot, so any
schema change - even a backward-compatible one - fails CI until the
snapshot is deliberately regenerated with
'pixi run update-schema-snapshot' in the same PR. The class list is
derived from LinkDef.h at configure time, so newly linked classes enter
the snapshot automatically; this also covers classes not exercised by
the reference files (detector wrappers, TrackFitResult).

Assisted-by: claude-code:claude-fable-5

The workspace build directory is the one the compatibility suite and
scripts/release.sh drive, but the configure task left BUILD_TESTING at
whatever the CMake cache happened to hold. A directory last configured the
way the conda package builds it (-DBUILD_TESTING=OFF) therefore contains no
build/tests, so pixi run test silently exercises nothing and the release
script cannot find build/tests/write_reference. Passing -D on the command
line also overwrites a stale cache entry, so the directory is brought back
into a testable state rather than needing to be wiped by hand.

Assisted-by: Claude Code:claude-opus-5

scripts/release.sh now writes tests/data/reference_v<version>.root from
exactly the tagged code, with the ROOT version pinned in pixi.lock at
that moment - capturing both the schema and the writing ROOT -
refreshes the schema snapshot, and gates the release on the test suite
(which picks up the new file via the CMake glob). Refuses to overwrite
an existing reference file.

The file is provisional until it has been validated: it must sit at its
frozen path while the tests run, so the EXIT trap covers that path too
and is only cleared once the snapshot regeneration and the test suite
have both passed. A release that fails validation therefore leaves
nothing behind for the exists-check to trip over on the retry, which
git checkout could not clean up on its own.

The writer runs through pixi like every other step, rather than relying
on the RPATH baked into the binary, and a missing build/tests explains
itself instead of failing on a bare "No such file".

Assisted-by: claude-code:claude-fable-5
Assisted-by: Claude Code:claude-opus-5

tests/data/README.md is the contract: which files exist, why the frozen
ones are never rewritten, the failure playbook for the compat and
snapshot tests, and the value recipe (language-neutral, so a future
Python/uproot reader can check the same files). README and CONTRIBUTING
point to it.

Assisted-by: claude-code:claude-fable-5