v0.95.0 — the ui:: namespace + Colossal Cave Adventure fully playable

Downloads: all six platform binaries for the v0.95 line are attached to v0.95.2.

madc v0.95.0

madc plays Colossal Cave Adventure. This release introduces the
ui:: data-hub namespace and proves it with a complete, fully playable
port of the classic game — a pure madc multi-file project whose output
is byte-identical to the original C open-adventure across all 94
reference transcripts — while the compiler itself got dramatically
faster to start and floating-point code got dramatically faster to run.
Everything since v0.92.1 (v0.93.0, v0.94.0, and v0.95.0) ships in the
attached packages for Linux (.deb/.rpm), Windows, and macOS
(Apple Silicon + Intel).

Features

• The ui:: namespace — worlds, projections, verbs, and prompts
  (documented in docs/language/ns-ui.md). A madc program opens a
  world file, reads and writes entity state through one hub-managed
  write path, renders text projections where projection selection IS
  the access decision, and dispatches verbs:
  long w = ui::world_open("adventure.world");
ui::act(out, w, player, "take", "lamp");
  Interaction is one verb too: ui::prompt(line, "> ") writes the
  prompt, flushes, and reads — piped transcripts read exactly like
  interactive sessions, which is how the game's logs stay
  byte-identical while a terminal player finally sees the > prompt.
• Colossal Cave Adventure, in madc, fully playable —
  examples/adventure/: 11 translation units built as one project
  (madc advent.cc.json), all state on a single class Game object,
  4447 lines against the reference's 4681 C lines. The 94-transcript
  parity corpus is a permanent test gate, so the game cannot silently
  regress.
• Write madc with zero ceremony — no #include, no using, no
  std::: bare print/println/format (including
  println(stderr, ...)), madc::getline(line) for input, quoted user
  modules auto-included, and array literals with a chainable push:
  var ds = { a, b, c }; ds.push(d);
• The compiler starts in the ~150 ms class on a real 11-file
  project (was 829 ms packed at the start of the arc; the dev binary
  fell from 1.9 s to ~0.3 s). The levers: auto-include that no longer
  matches std names in impossible positions, lazy first-call codegen,
  one shared compile substrate and shared prelude images across a
  project's translation units, lazy template hydration (a C-shaped
  launch never decodes the 6.4 MB template payload), a de-RTTI'd hot
  front end, and page-arena allocation in c2mir.
• Floating-point code runs ~2.8× faster — donut.c now beats
  gcc -O0 (v0.93.0): MIR's scalar SSE converts carried a false
  dependency that serialized call-heavy FP loops; the fix is the same
  pxor dependency-breaking idiom gcc and clang emit.
• Every polyglot function has a lean primary form — the 57
  PHP/Python/Perl/Ruby/JS/Rust functions that only accepted
  std::string now take value/char* directly with the source
  language's real semantics (perl::chop mutates — that IS Perl;
  PHP string functions return new strings).
• Sturdier codegen, from the community (v0.94.0): a wrong-code
  -O2/-O3 ssa_combine loop-PHI fold fixed (reported with reducers
  by ThePeiLin, upstream mir#467 — our fix, submitted upstream),
  the register-allocator spill-table bound adopted from
  Bill Hlavacek (upstream PR #468), and the aarch64 displacement
  gate adopted from Richard Davison (upstream PR #466) — the last
  directly speeding madc's Apple Silicon lane.

Fixes and internals (compact)

• Cross-TU correctness under --project, each with a pinned test:
  every TU's global initializers now run (was: last module won);
  extern var no longer creates a private per-TU copy; file-scope
  member defaults (NSDMI) apply; --no-embedded-headers reaches
  project TUs (a new fulltest gate covers the whole CLI policy-fanout
  family).
• A call with no viable overload refuses loudly instead of silently
  binding the wrong function; carrier subscripts are true lvalues over
  the element slot (arr[i] = x writes land); frozen-value container
  mutations degrade to a diagnostic instead of aborting.
• Duplication audit at the merge wave: one MEMBER param-table decoder,
  forest decode accounting restored, the last zero-filling decode
  destinations converted, remaining families recorded with markers.
• Build-infrastructure hardening: the container build config no longer
  crosses the sync tunnel (two masked staleness traps closed), and
  bin/madc-release — the profiling vehicle — is rebuilt as a standing
  lane of every merge wave.

See CHANGELOG.md for the complete per-release
record (v0.93.0, v0.94.0, v0.95.0).