weir v0.0.81

Fixed

• Launch is faster for scripts that import large modules. The
  check-time walk that decides which module bindings are command
  helpers asked, for every variable it met, whether the name was a
  reifier's desugar spelling, and built 28 strings to answer each time.
  The answer is now one table lookup. A script importing the 6,400-line
  starship port goes from about 750 ms to about 400 ms before its first
  line runs; bare startup was and is about 12 ms.
• The check itself is leaner on large modules. The module loader
  re-derived its command-helper table from every earlier binding on
  each statement; the assembler, the district mask and doc attachment
  each rescanned every physical line for its comment several times;
  the load-target scan re-parsed the whole file; constructor lookup
  rebuilt the set of imported types per use; and the command grammar
  rebuilt its resolver-independent parsers on every attempt. Each now
  does its work once. Same diagnostics, same trees: the starship import
  goes from about 400 ms to about 250 ms (CPU-pinned medians 373 → 254
  ms), a direct check of its largest file from 147 ms to 95 ms; bare
  startup stays about 12 ms.
• Imported modules check in parallel. The import graph is read
  before any module checks, and a module whose imports are done checks
  on a worker thread (as many as processors, at most 8) while the
  importing file's check walks toward its import line; each result
  surfaces at that line in load order, so diagnostics are unchanged to
  the byte (an 80-file parity corpus, the starship port checked 50
  times). On a 15-core box the starship import goes from 263 ms to 215
  ms (medians of 9); pinned to one core it is the sequential loader,
  249 ms as before; bare startup stays about 12 ms. What remains is the
  collector's: the workstation GC marks on one thread, and the module
  checks spend most of their parallel phase stopped for it.
• The collector has less to do. Merging an imported module into the
  importer rebuilt its set of taken type names once per type and walked
  every type origin through an enumerator per lookup; it now keeps both
  as it goes. Finalizing a statement's types and typed tree copied every
  node even where nothing changed, and renaming an import's types copied
  ones no rename touched; unchanged nodes are now shared. Same
  diagnostics, same trees: the starship import allocates 254 MB instead
  of 287 MB, promotes 13.3 MB instead of 15.7 MB out of the youngest
  generation, and spends about 50 ms instead of 57 ms in GC pauses; it
  goes from 246 ms to 225 ms pinned to one core and from 206 ms to 184
  ms on 15 (medians of 25), a direct check of its largest file from
  91 ms to 86 ms; bare startup stays about 10 ms.
• The parser builds its grammar once. The command grammar was
  rebuilt for every statement, and for every command line nested in a
  let, an if or a value-headed pipe, because it held that statement's
  name resolver; it now asks the current one as it parses. A postfix
  chain rebuilt its index, slice and field parsers at every step, the
  expression grammar's teaching alternatives read the source position
  before checking the next character, and every node's span allocated
  two positions on the way. Same diagnostics, same trees: the starship
  import's parse allocates 79 MB instead of 151 MB, and the import goes
  from 225 ms to 202 ms pinned to one core and from 192 ms to 167 ms on
  15 (medians of 21); bare startup stays about 9 ms pinned.
• A misspelled flag no longer prints the next argument. After an
  unknown flag, Args.load reported the following word as an
  unexpected argument and quoted it, so --tokn hunter2 for a
  Secret flag printed the secret. The word is still reported, as
  unexpected argument after '--tokn', without its text.
• cmd | exec honours the current directory, and is refused where
  it would skip cleanup. The replaced process started in the
  script's start directory whatever cd or within cd said; it now
  starts in the session's. Inside any within block (tmp, proc,
  lock, serve, cd, env, or a bare within … always) exec is
  now a check error: replacing the process left the temp dir behind,
  orphaned the scoped process and never ran always.
• let x = retry … runs each body line on its own. With the loop
  head on the let line, same-indent body lines were glued into one
  command (echo first echo second) with only a warning; they now
  sequence as they do without the let.
• An explicit Content-Type replaces the body's. A request with
  a Content-Type header and a Text/Json/Form body sent both,
  comma-joined (text/plain, text/csv), which servers reject. The
  explicit header now wins and exactly one is sent, as for
  User-Agent.
• #save keeps what it promises. A line that used a name later
  redeclared was saved above the redeclaration and failed at run, and
  an alias over a builtin (#alias ls = ls --color=never) dropped the
  aliased let while keeping its dependents. Both saved files now
  check and run.
• A top-level YAML null writes null. to yaml wrote empty
  text for it, which does not read back, and a null document
  vanished from a to yaml stream.
• Server.port reports the bound port. See Changed.
• Dir.move crosses filesystems. A move to another mount failed
  with "cannot access … Invalid cross-device link"; it now copies then
  deletes, as mv and File.move do, and its errors name the paths.
  A symlinked source moves as the link, exactly as on one filesystem.
• weir restore repairs a locally modified signature. It said
  present and exited 0 while verify said MODIFIED.
• Instant.tryParseWith raises for a format without a full date.
  Such a format can never match, so returning None hid the bug; it
  now raises as parseWith and an unknown directive do.
• Http.withQuery puts the query before a #fragment. It
  appended after it (h/p#frag?a=1), so the parameter was never sent.
• [<Wire "name">] matches a table header verbatim even when it
  is spelled like the field.
• /// above a one-line [<Attr>] field: T reaches --help.
  The own-line attribute form worked; the inline one dropped the doc.
• weir -e and the REPL echo share one renderer. Piped, both
  write a value's control characters byte-faithfully; at a terminal,
  both show them visibly (\x00, \x1b). They used to disagree.
• No more .NET exception text from show of the minimum
  Duration, strings past the length limit (Str.join, +,
  interpolation, Str.replace, Bytes.toHex now name the limit),
  Json.inferShape on bad JSON (and the message names
  Json.inferShape, not #infer), an init file's bad #session env
  key (now a check error, NUL included), an out-of-range serve port,
  an Other method that is not a token, an invalid or non-http URL, a
  Transfer-Encoding request header (refused before send), a garbled
  status line, weir add schema given a local path (add schema takes an http(s) URL), and a malformed .weir/lock.json.
• weir check agrees with the run. It refused spec-valid code —
  let q = 3 in q + 1 inside an indented block, cmd | and cd d in a
  block — and passed code the run refused: Seq.pmap/piter/
  pfirst/pfor and a mutating Http.expect inside plan,
  [<Short "-">] (a short is now one letter or digit). A misplaced
  or unknown directive gets the run's message from check too, an
  ambiguous imported case is named as such, env-key refusals are
  located where the key is built, and a shared module's error is
  reported once, not once per import.
• weir fmt no longer refuses a match arm continued by |> under
  its body, a layout the checker accepts.
• Clearer messages. No internal names ('__hole1, 'a1) or
  parser expectation dumps in diagnostics; one shape for a bad flag or
  env value (--rate: not a float: 'abc'); -- is refused as an
  end-of-options marker instead of getting a did-you-mean; --help
  keeps two spaces between a long flag and its text (wrapping when it
  must) and shows string defaults quoted; a TLS handshake with a plain
  HTTP port is no longer called an untrusted certificate; and a dozen
  more located or reworded errors.
• Docs match the binary. The reference and the skill file no
  longer contradict behaviour: Seq.groupBy forces its source,
  Seq.take past the end truncates while skip raises, Size's
  show does not parse back, identifiers take no ', from yaml
  refuses Map fields, a yaml patch's sequence appends depend on
  patch order, File.exists is false for a directory, command output
  starting with a UTF-16/32 byte-order mark decodes as that encoding,
  and the echo shows 100 elements in -e and piped as at a terminal.

Changed

• ServerConfig.port is an Option<int>. Some 8080 binds that
  port; None asks the kernel for a free one, and Server.port
  reports the port actually bound (it said 0 for port = 0). Migrate
  { port = 8080; … } to { port = Some 8080; … }.
• <, >, <= and >= compare strings, bools and Options.
  Whatever Seq.sort orders, the operators now compare (strings
  ordinally, false before true, None first); Uuid still has no
  <.
• The -e and piped REPL echo shows 100 elements of a lazy seq,
  as the terminal REPL does and #echo reports (it was 10).