[ianhi]

Adds a way to answer format-level questions about an Icechunk repo by reading its metadata files (repo info, snapshots, manifests, transaction logs) field by field, from the CLI, MCP, or a browser. Motivating questions from a user:

1. Which chunks changed between two versions?
2. How does a manifest record a chunk deletion?
3. What checksum does a virtual ref store for its target object?

What's new

Command	What it does
core-drill <repo> object <target>	Decode one metadata file. IDs that name other files print as → manifests/<id> links; node IDs get their paths; compressed_location is decompressed; --at arrays/0/refs/100..150 selects a subtree or window.
core-drill <repo> chunk-ref <array> <coords>	Find the manifest entry for one chunk: native chunk ID, inline bytes, or virtual location with its ETag / last-modified checksum.
core-drill <repo> chunk-changes <commit>	Classify each coordinate in a commit's transaction log as added, overwritten, rewritten, deleted, or absent, by looking it up in the parent's and the commit's manifests.
core-drill <repo> web	Serve the viewer from the binary on 127.0.0.1 and open it on this repo. Repo objects go through core-drill's storage connection, so S3, GCS, and Arraylake repos work without CORS.

All three lookups are also MCP tools (object, chunk_ref, chunk_changes) and support --output json.

$ core-drill ./repo chunk-changes MNB5ZD7ZTTZT01TTXV6G
## /native — 1 coordinate listed
- (1, 1) **deleted**
  - before: native chunk `P7TN6M0SRMNJMNXGXJ80`, 4318 bytes at offset 0 — in `manifests/NC0XXACBQRMJMMHKFRF0` at `arrays/0/refs/3`

How it works

• Schema-driven decoding. schema/*.bfbs is compiled from the Icechunk .fbs files (scripts/gen-schema.sh). The Rust walker (src/raw/) and the JS viewer (web/decoder.js) both read every field by reflection, so a schema change only needs a regenerated .bfbs.
• Untrusted input. Reads are bounds-checked, and verification goes through icechunk's own constructors. A corruption test truncates and flips every byte of each file type. flatbuffers-reflection's verifier is not used because it panics on out-of-range union tags. FlexBuffers decoding is bounded and panic-contained, and sanitize now also strips bidi control characters.
• The viewer is self-contained: no build step, the zstd wasm decoder is vendored (web/vendor/zstd.js), and there are no CDN requests. web/test/decode.test.mjs checks it field for field against core-drill --output json.

Also in this PR

• storage-size command and MCP tool: deduplicated on-disk bytes across history.
• Output subcommands default to markdown instead of launching the TUI.
• Upgrade to icechunk 2.2.2 and arraylake 0.14, which fixes opening Arraylake marketplace subscription repos ("Bucket info is missing"). arraylake 0.14 resolves a subscription to its parent repo's bucket.
• cargo fmt over the repo (separate commit).
• --locked in the from-source install command.

Tested with:

• 37 Rust tests, including end-to-end lifecycle and corruption tests on an in-memory repo.
• 133 parity checks between the web decoder and the CLI across four local test repos (node web/test/decode.test.mjs).
• Manual runs against s3://icechunk-public-data/v1/era5_weatherbench2, al:iansorg/icechunk-format-examples (generated by scripts/create_chunk_lifecycle_repo.py --arraylake …), and a marketplace subscription repo.

Docs: DOCS/raw-inspection.md, web/README.md.