Co-Authored-By: Claude Code <noreply@anthropic.com>

Candidate PR: replace the ReInterHand `shared-v1.0` tree with the verified
`shared-v1.1` release, as in every other dataset's v1.1 PR. **Please review before
merging.**

Full local migration of 6,750 views across 500 sequences (6,250 exo +
500 ego cameras, 1,526,795 frames), validated and audited locally: v1.1 merges each group
of numeric tensors and its `valid` into one `.pt` per view (`camera/camera.pt`,
`annotation/mano/mano.pt`) and leaves media, JSONL and the rest of the metadata alone. This
PR adds `ReInterHand/shared-v1.1` (54,641 files) and removes
`ReInterHand/shared-v1.0`; it does **not** touch `ReInterHand/raw`, and no other dataset's
files or LFS tracking lines are modified (both re-checked on the candidate branch before
merging).

Pushed under yhl's authorization; merging is left to the reviewer.

Why this PR removes `shared-v1.0`:

- The repository is at the Hub's 1,000,000-file ceiling, and a push that would cross it is
refused outright -- the ARCTIC candidate PR (#5) was answered with "Your git repo would
contain 1000002 files after this push, over the limit of 1000000 files. Offending
reference: refs/pr/5". Removing the old tree is what makes room for the new one: this PR
adds 54,641 paths and takes the old tree's paths out on the same candidate
branch, so the repository ends up smaller than it started.
- Every other v1.1 PR does the same: dex-ycb is already on `main` that way, PR #4 does it
for HOT3D / OakInk-v1 / OakInk-v2 / FreiHAND, and the ARCTIC and RHD candidate PRs do it
too. This release had registered itself as the exception -- its own
`doc/PUBLICATION_PLAN.md` records the duplicate tree as a deviation from the registry's
`release_layout`, with the removal deferred to a separate change once the publication
manifest and every path, size and content digest had verified. That is this PR; the user
authorized the removal on 2026-10-15.
- Nothing unique is destroyed. The published v1.0 `doc/validation/` -- the iipl_100 review
record the plan had counted as existing nowhere else, 376 files plus 5 aggregates that
describe 500 sequences rather than this machine's 272 -- is Hub-side only: iipl_100 still
has it under its own `shared-v1.0/doc/validation/`, and the published tree stays
fetchable at the pre-merge revision after the merge. This working copy's own
`shared-v1.0/` is untouched; only the Hub's copy goes.

Conventions this release carries deliberately:

- Media is **not re-uploaded**. All 13,500 media streams (one `rgb.mkv` and
one `mask.mkv` per view, ~1.08 TB) are byte-identical to the published v1.0 content and
are not re-encoded, so 13,445 of them go up as server-side copies from
`ReInterHand/shared-v1.0` at the same relative path (copies resolve against `main`, so
the removal commit cannot break them) -- including the 13,446
this working copy does not have at all (`MIGRATION.md` §3/§6 keeps unchanged media where
it is, and that absence is declared in the tree-external
`tmp/reinterhand-tools/media-presence.json`, which is not part of the release). The one
exception is 1 stream (89,668 B) that the Hub stores as a plain
git object rather than an LFS object: no copy source exists for it, so its bytes were
fetched from the published v1.0, checked against that file's size and git blob id, and
uploaded. It is the only media byte that travels.
- Storage, in total: 27,113 of the 54,641 files are
server-side copies and 27,528 are uploaded (1.75 GiB of new
bytes), so the release goes up without reading 1.08 TB of video.
- File counts, since the point of this migration is fewer, larger files: a view goes from
12 files to 8 (exo) or 14 to 8 (ego) -- the numeric tensors merge into two `.pt` per
view -- which is why the tree is 54,641 files.
- `doc/validation/` and `doc/history-v1.0/`: the release ships 27 per-view visual review
pages plus `V1_1_VALIDATION.json` (acceptance evidence, not training input; H2O, dex-ycb,
ARCTIC and RHD ship the same kind of pages), and reproduces v1.0's `doc/` (752 MB,
including 114 rendered review clips) byte for byte as `doc/history-v1.0/` -- a history
tree whose pages link to videos left behind would not be a reproduction. Because this
machine's v1.0 is the iipl_126 half plus iipl_100's `data/`, that reproduction and
`doc/V1_0_SEMANTICS.md` are of *this machine's* v1.0, and its `doc/validation/` covers 272
sequences rather than 500; `data/` is unaffected -- its 68,554 files were verified against
the Hub file by file. The published v1.0 `doc/`/`script/`, including the iipl_100 review
records this machine never had, are removed with the rest of the old tree; this release
keeps a copy of *this machine's* v1.0 `doc/` as `doc/history-v1.0/`, and
`doc/PUBLICATION_PLAN.md` records both facts.
- `.gitattributes`: this repo tracks LFS with a mix of global patterns (`*.pt`, `*.png`,
...) and explicit per-path lines, and at `main` that file is 9,580,405 B against the Hub's
10 MiB ceiling for non-LFS files -- 905,355 B of headroom. `.mkv` has no
global pattern, so the Hub appends one line per new `.mkv` path: this release's
13,446 of them would need ~2.7 MB and cannot fit. ReInterHand's own 13,007
lines in that file are exactly its v1.0 `.mkv` paths, and are 2,636,472 B. This PR
replaces them with the glob form the repository already carries for
`dex-ycb/shared-v1.1/**/*.obj` (live on `main`), for HOT3D / OakInk-v1 / OakInk-v2 /
FreiHAND (PR #4) and for ARCTIC and RHD: the file becomes 6,944,071 B,
3,541,689 B of headroom, and the Hub then appends no per-path lines for this
release. Every non-ReInterHand line is byte-identical, re-checked on the candidate branch
before merging.
- The nine media-component metadata gaps that v1.0 carried (`storage.profile`
`mkv/ffv1_rgb24/v1` is unregistered and outside `rgb/uint8/v1.allowed_storage_profiles`;
the `segment` component has no `storage` block and lacks `segmentation_type` /
`coverage` / `entity_id`) are **not** fixed here: the v1.1 contract is a pure storage
migration, the media components are byte-identical between the two versions, and fixing
them is a separate, intentional change. They are declared and counted in
`script/reinterhand_v1_1/reinterhand_common.py` (`DECLARED_MEDIA_SCHEMA_GAPS`, exactly
6,750 views each), and `doc/PUBLICATION_PLAN.md` records them.

The release's own status flags (`upload_authorized: false` in all 6,750 views,
`full_output_human_review: "pending"`, and `human_full_review: pending`,
`publication_authorized: false`, `hf_writes: false` in `doc/FULL_MIGRATION_SUMMARY.json`)
are untouched: they describe human review state, which is recorded outside the tree. After
this PR merges, the full-tree validation (`--expect-views 6750`, without
`--media-declaration`) still has to be run on the complete tree, which this machine cannot
do.

🤖 Generated with [Claude Code](https://claude.com/claude-code)