Version 0.9.0 - POSIX conformance achieved

posixutils-rs v0.9.0

Depth over breadth: a POSIX.1-2024 conformance pass across the tree

1,009 commits (70 PRs) since v0.8.0 (2026-06-25); 817 files changed,
+178,348 / −66,622.

Nearly every utility touched this cycle was audited against POSIX.1-2024
clause by clause, and the findings fixed. The C compiler alone accounts for a
third of the commits.

Highlights

• pcc is now c17 — the C compiler driver ships under the name of the
  one language it compiles. Dialect switching is gone; -std= remains only as
  a compatibility shim.
• make builds real Makefiles — VPATH, built-in inference rules, %
  pattern rules, $(eval), the macro function set, conditionals and
  define/endef.
• pax speaks tar and cpio — both command lines route into pax.
• man renders pages real systems ship — the pest grammar is gone,
  replaced by a roff/mdoc interpreter with tbl text blocks, man-db config
  support and in-process decompression.
• Byte-faithful text handling — sh, tr, diff, echo, expr, pax,
  i18n and plib carry file and argument data as bytes, so a non-UTF-8
  filename or IFS=é no longer aborts anything.
• Dependency diet — the workspace went from 159 to 102 crates.
• aarch64 Linux CI — a second architecture now gates every push.

C compiler (c17, cflow, ctags, cxref)

• Renamed pcc → c17; compiles exactly one language, C17.
• ABI correctness across x86-64 and AArch64: HFAs, all-SSE and mixed structs,
  MEMORY-class aggregates, stacked-argument placement and alignment,
  __int128 and binary128 pairs, Darwin variadic aggregates, and the
  register tally a fixed parameter actually spends.
• New types and extensions: __float128 (IEEE binary128), _Float16,
  _Generic, __attribute__((mode)), transparent_union, vector_size,
  computed goto, case ranges, designated-initializer ranges, GNU ?:,
  __real__/__imag__, #pragma pack, _Atomic operators that are actually
  atomic.
• Builtins: checked arithmetic, floating classification, __builtin_va_arg_pack,
  __builtin_object_size, the fortified _chk entry points, the ffs/clrsb
  family, and a working <tgmath.h>, <stdint.h> and <stdatomic.h>.
• Preprocessor: __VA_OPT__, -dM, -P, -include, the -M dependency
  family, #include_next, include-guard detection, a correct macro hide set,
  and -E output that keeps its source lines.
• TLS: dynamic model with TLS descriptors; -fPIC selects a
  position-independent model.
• Debug info: -g now emits line information a debugger can use, plus DWARF
  descriptions of functions and their variables.
• Constant folding rebuilt to fold each constant at its own width and in the
  expression's own format.
• Diagnostics: dozens of constraint violations that used to compile silently
  are now rejected, and diagnostic message bodies are translatable.
• cflow analyzes object files, assembler source, and lex/yacc input;
  ctags/cxref output formats corrected.
• CPython 3.12.9 builds and passes its full test suite at -O0 and -O2.

Shell (sh)

• Expansion, lexing and pattern matching carry byte strings end to end.
• Field splitting rewritten as one state machine transcribing POSIX 2.6.5.
• Job control: async lists, bg/fg, blocking waitpid, status retention
  until wait collects it, and refusal of fg/bg in a subshell.
• Built-ins: test/[ sharing one evaluator with the utility, pwd, true,
  false, cd -e, a conforming ulimit, umask, unalias, fc.
• Correct exit statuses, $-, $!, errexit, and-or short-circuiting,
  subshell traps, and command-substitution status.
• Fixed two here-document hangs, eight process-aborting panics, and pipeline
  descriptor leaks.
• New PTY test harness.

Editors (vi, ex, ed)

• vi: :map/:unmap/:ab/:una with shortest-match lookup, a tag stack
  (:tag, ^], ^T), autoindent with spec-correct ^D, abbreviations, and
  key input driven through a real terminal queue.
• vi: undo recorded for every buffer-changing command; motions classified so
  operators act on the right region; character and display-column counting
  where bytes were counted; buffer preserved on hangup.
• ex: address resolution rebuilt (separators, marks, line 0, excess
  addresses), POSIX write and join rules, real substitute flags and
  replacement escapes, +command and argument lists.
• ed: mark tracking across structural edits, one substitute per undo, wq
  parsing, n output format.

Build tools (make, lex, yacc, nm, strip, ar)

• make: VPATH and the vpath directive, built-in inference rules, %
  pattern rules, $(eval), macro functions, conditionals, define/endef,
  override ranking, .SCCS_GET retrieval, PROJECTDIR search, -p database
  dump in makefile syntax, and a ledger-based build state. The rowan CST was
  replaced by a line scanner.
• lex: equivalence classes derived from the NFA, per-start-condition entry
  states, ^ resolved by automaton rather than post-match filter, REJECT
  history, bracket-expression conformance, YY_INPUT.
• yacc: LALR closure termination, non-productive non-terminal rejection,
  precedence from the last terminal, $n resolution in mid-rule actions,
  yytname in internal symbol order, $undefined.
• nm/strip: filter interaction, section symbols, long archive member names.

Archives and file transfer (pax, compress)

• tar and cpio front-ends.
• Extraction is fd-relative, so a planted symlink cannot redirect a write.
• Extended headers: path= records for oversized names, non-UTF-8 pathnames
  under -o invalid=binary, -o keyword options when listing, time keywords.
• Correct -s substitution, -u per member, hard-link contents in cpio,
  multi-volume ustar prefixes, and attribute restoration on copy.
• compress preserves mtime at nanosecond resolution and pre-1970 timestamps.

Text utilities

• diff: -c/-u built from grouped sections, byte comparison, directory
  differences in the exit status.
• patch: reads context diffs the way diff writes them, refuses absolute
  and parent-directory names, detects format over the whole patch, recognizes
  mailed headers.
• tr: locale-driven case conversion and class membership, -c vs -C,
  equivalence classes, the string2 class rules.
• od: float rendering as shortest-round-trip %g in aligned columns, -t f
  size letters, long double conversion, block-group duplicate suppression,
  -N as a byte count.
• printf: XBD Ch. 5 precision and flag rules, %n$ numbered conversions,
  unsigned conversions that saturate, incremental writes.
• more: streaming input, cell-based measurement, LINES/COLUMNS on resize,
  and a PTY test harness.
• split, csplit, strings, uniq, pr, expr, echo, find, du,
  getconf: conformance and diagnostic fixes throughout.

Other utilities

• mailx: one command table in place of three interpreters, mailbox locking,
  mbox that reads back, blind recipients, terminal-derived screenfuls.
• talk/talkd: the local transport works at all now; terminfo-driven
  drawing, BSD address precedence, signal-safe teardown, end-to-end tests.
• sccs (ten utilities): one shared implementation of each rule, s-file
  protection enforcement, delta by minimal diff (and on binaries), -i/-x
  SID lists, z-file locking with registered cleanup.
• cron/crontab/at/batch: CRON_ALLOW/CRON_DENY gating, signal
  handlers installed before forking, $SHELL as the job interpreter, edits
  preserved on editor failure.
• bc: pest grammar replaced by a hand-written lexer and parser; scale-correct
  division, sqrt and modulus; POSIX dynamic scoping; streamed, column-wrapped
  output.
• m4: GNU differential gate, lazy -s line directives matching GNU byte for
  byte, ifelse warning parity.
• man: see Highlights; also -M, -S, -k/-f native NAME search,
  architecture subdirectories, .so restricted to manual roots, recursion
  bounds throughout the roff interpreter.

Internationalization

• gettext-rs loads real message catalogs, honoring NLSPATH, with
  positional-argument and plural formatting.
• Diagnostics across the tree route through plib::diag / io_error_text, so
  no utility leaks Rust's Display or (os error N).
• Locale-aware collation, case mapping and character classes in plib.

Library and infrastructure

• plib: byte-oriented regex and collation, plib::tmp (replacing tempfile
  and retiring rustix), plib::tty with the terminal-write TOCTOU closed,
  a descriptor guard, close-on-exec temporaries, and strerror-faithful
  error text.
• Dependency reduction from 159 to 102 crates; one codegen unit per crate for
  release builds.
• CI runs the suite on aarch64 Linux as well as x86-64 Linux and macOS.
• Audit files across the tree were re-probed by execution and reduced to their
  open items; cflow, ctags, cxref, echo, printf, talk, talkd,
  tr and csplit were promoted to Stage 6 (Audited).

Security-relevant fixes

• pax extracts fd-relative; patch refuses absolute and parent-directory
  paths.
• gettext-rs bounds catalog allocation taken from an untrusted header.
• cron/lp guard environment overrides against set-gid, not only set-uid.
• Temporary files are created close-on-exec, with cleanup armed and identity
  logged.
• Recursion and output budgets bounded in man/roff, m4, bc, make and
  lex.

Contributors

Jeff Garzik, grisenti