Add a core regression test for MosaicReader::page_infos asserting
plain/dict/const detection on a paged-bucket file, and CLI unit tests
for the fmt helpers (json escaping, value/encoding rendering, ndjson
null handling, table truncation).

Drive the mosaic binary against a fixture file (via CARGO_BIN_EXE) and
assert stdout for schema/meta/pages/cat, --json output, projection,
row truncation and missing-file failure. No external dev-deps.

Align the viewer command set with parquet-cli/arrow-rs: head (alias of
cat), footer (magic/version/buckets/compression), column-size (on-disk
bytes per column), dictionary (dump dict-encoded entries). Core gains
compression()/dict_values()/dictionary() read-only accessors. e2e tests
cover the new commands.

Mosaic's column-bucket grouping has no parquet equivalent. Add a
buckets command printing, per row group, each bucket's kind
(empty/monolithic/paged), on-disk size and member columns. Core gains
MosaicReader::bucket_infos(). e2e covered.

Align dictionary column selection with parquet-cli's -c flag instead of
a positional argument; update e2e.

Completes JSON output across all 9 commands; dict columns emit an array,
non-dict row groups emit null. e2e extended.

Expand docs/cli.html and cli/README.md to cover every command
(schema/meta/footer/buckets/pages/dictionary/column-size/cat/head) with
usage and example output. Drop all comparison content per maintainer
preference.

Remove the near-trivial encoding_names mapping test; extend footer and
buckets e2e to cover their --json output, improving CLI feature coverage.

The e2e tests carry their own fixture writer; the standalone gen.rs
example duplicated it and was unreferenced.

BucketInfo now carries uncompressed size (exact for monolithic, unknown for
paged). column-size prints a total with uncompressed bytes + ratio; buckets
shows per-bucket ratio when known. Tests + docs updated.

count prints total rows; --all drops the -n cap; --where applies one
column/op/value condition (=,!=,>,>=,<,<=), numeric or string. Tests + docs.

convert imports a CSV into a new Mosaic file (schema inferred, --stats picks
min/max columns). cat --where now skips row groups whose min/max provably
exclude the predicate, conservative (any missing stat keeps the group).
Tests + docs.

convert dispatches on extension: .json/.ndjson/.jsonl read one object per line
via arrow-json (CSV path unchanged). Test + docs.

… atomicity

- page_infos validates paged slot sizes sum to total (rejects forged sizes
  that could drive a ~4GiB read, matching the projected reader)
- cat --where: ordering ops on a non-numeric column/value error instead of
  silently dropping all rows
- cat --json: NaN/Infinity emit null (was invalid JSON)
- convert writes a temp file and renames on success (no truncated output)

… column-size

- cat --json renders via Arrow JSON writer: all reader types valid, NaN->null
- where: numeric cols compare numerically, others exact string; ordering numeric-only
- column-size: paged uses exact slot sizes, monolithic multi-col marked (approx)
- README: scope --json to inspection/query commands (convert writes a file)

- where and stats_exclude compare integers in i128 (full i64 range exact);
  large ids (Snowflake) no longer mis-skip or mis-match via lossy f64
- ndjson returns io::Result instead of .expect, keeping the CLI error contract

…red 0.3)

Replace arrow-array/schema/select/csv/json with arrow = {features=csv,json};
one version to bump, paths via arrow::{array,datatypes,error,compute,csv,json}.

Read the filter column even when projected out, then drop it before printing,
so cat -c id,name --where score>440 filters instead of erroring.