v0.182.0

Upgrading: nothing to do (INDEX_VERSION stays at 124). Only snapshot inspect and snapshot create --out change: for the file names SQLite reads
as something other than a file (file:…, :memory:, the empty name), in
SQLite's error text for a relative name, and for a SQLite header whose bytes
18 and 19 disagree.
To pin the binary back: npm i -g @sdsrs/code-graph@0.181.0
(or, from source, cargo install --locked --git https://github.com/sdsrss/code-graph-mcp --tag v0.181.0 --features embed-model), then delete ~/.cache/code-graph/bin/code-graph-mcp and set
CODE_GRAPH_NO_AUTO_UPDATE=1, as described under 0.171.0.

Measured on 28 inputs, the 0.181.0 release binary against this one. Thirteen
changed; the other fifteen print, exit and leave the same files on both.

A file name SQLite reads as something else names that file

SQLite reads three kinds of name as something other than a file: :memory:
(a database in memory), the empty name (a temporary database), and, because
the SQLite built into this binary accepts URIs whatever the caller asks for,
any name that begins with file:, so file:x.db meant x.db. Where the two
commands gave SQLite a relative name as typed, they now put ./ in front of
it, which makes each of these a file in the current directory. An absolute
name is passed as before.

• snapshot inspect file:x.db read x.db. With no x.db beside it, it
  exited 1 with unable to open database file: file:x.db; with another
  snapshot at x.db it printed that snapshot's metadata (exit 0); with x.db
  in WAL mode and a -wal beside it, it printed the value in that -wal and
  left an x.db-shm behind. It now prints file:x.db's metadata in all three
  (exit 0) and leaves no file behind. Three inputs.
• snapshot inspect :memory:, on a snapshot named :memory:, read an empty
  database in memory and refused it as not a valid code-graph snapshot
  (exit 1). It now prints the snapshot (exit 0). One input.
• snapshot create --out file:x.db wrote x.db (exit 0). Where x.db
  already held a database it exited 1 (VACUUM INTO 'file:x.db', output file already exists) and left x.db as it was; --out file:sub/x.db beside a
  directory file:sub exited 1 (unable to open database: file:sub/x.db).
  All three now write the file they were given (exit 0). Where file:x.db
  itself already held a database, create wrote x.db and exited 0; it now
  exits 1 with output file already exists, as --out x.db does when x.db
  holds one. Four inputs.
• snapshot create --out :memory: exited 0 having written nothing; it now
  writes a file named :memory: (exit 0). --out "" exited 0 having written
  nothing; it now exits 1 (VACUUM INTO '', unable to open database: ./),
  after the index has been built. Two inputs.
• SQLite's errors quote the name it was given, so for a relative name they
  now show ./ in front of it. Two inputs, exit 1 on both versions: inspect s.db, whose absolute path (730 bytes) is longer than SQLite accepts, said
  unable to open database file: s.db and now says ./s.db; create --out ro/x.db into a read-only directory said unable to open database: ro/x.db
  and now says ./ro/x.db, under the same VACUUM INTO 'ro/x.db' line.
• Unchanged, twelve inputs. Given to inspect: ./file:x.db, the absolute
  path, a file:x.db in WAL mode, a compressed file:y.db.zst, and a relative
  name in a directory whose name is not UTF-8. Given to create --out:
  ./file:x.db, file:x.db.zst, a relative name in a directory whose name is
  not UTF-8 (with and without a sibling directory whose name is its U+FFFD
  spelling), and a file already at --out that holds a database, is empty, or
  is a .db.zst.

A header that says WAL in byte 19 leaves no side files

SQLite opens a database's WAL when byte 19 of its header is 2; inspect
looked at byte 18. SQLite writes the two bytes together, so only a header
edited by hand splits them. Bytes 18/19 = 1/2 made inspect leave a -wal
and a -shm beside the file; it now leaves neither, and prints what it
printed (exit 0). Headers 2/1, 1/1 and 2/2 print and leave what they did. Four
inputs, one changed.

Not covered

• Of the items under 0.181.0's Not covered, the file: name and the 1/2
  header are fixed above. The other items listed there are not addressed by
  this release.
• create --out over an existing file, as in 0.181.0: a raw --out that
  holds a database is refused (output file already exists), an empty one is
  overwritten, and a compressed --out is replaced and gets a .blake3 beside
  it.
• --out "" is refused only after the whole index has been built.
• inspect cannot read a raw snapshot in rollback mode whose absolute path
  is longer than 504 bytes, as in 0.181.0: SQLite refuses a path whose
  -journal would not fit in its 512 bytes. Measured at 504 bytes (read),
  505 and 730 (refused), named relatively. A snapshot in WAL mode at 730 bytes
  reads, because it is read through a copy.
• create --out x.db.zst fails when the path of TMPDIR is not valid UTF-8
  (VACUUM INTO '<TMPDIR>/…/compacted.db', exit 1), as in 0.181.0. A raw
  --out x.db under the same TMPDIR is written.
• On Windows, : in a file name separates a file from one of its NTFS
  streams. create --out file:x.db wrote x.db there; it now hands SQLite
  .\file:x.db, which names a stream x.db of a file called file. --out :memory: reaches SQLite as .\:memory:, a name Windows cannot hold as a
  file, so the file named :memory: above is what Linux writes. Read from the
  Rust and SQLite sources, not run: there is no Windows machine here.