v1.1.0

Added

• A coverage report. nix build .#coverage (or sbcl --script run-coverage.lisp outside Nix) runs the regression suite under SB-COVER
  and writes an HTML report instrumenting cl-prolog and cl-prolog/weave
  only, not the cl-weave harness driving them. Nix embeds its runner in
  flake.nix with pkgs.writeText; run-coverage.lisp remains the local
  SBCL entry point. checks.coverage builds the Nix report in CI; it asserts
  the report exists, not a coverage percentage. See
  Testing.
• benchmarks/performance.lisp gains three benchmark groups matching the
  optimizations below: indexed substitution, ground-vs-nonground tabled
  answer replay, and the left-recursion cache across cold/warm/revision-
  changed rulebases.

Changed

• Fixed-builtin goal dispatch (%goal-solver) uses a nested eq (predicate)
  -> eql (arity) hash-table pair instead of one table keyed by a consed
  (predicate . arity) under equal, avoiding a cons and an equal hash on
  every builtin lookup — the hottest path in goal dispatch.
• Left-recursion detection caches its per-revision, per-module SCC index and
  its recursive predicate/arity membership check through nested eq/eql
  hash tables instead of an equal-keyed table over a freshly consed
  (rulebase revision module) list, and now reads the rulebase's own
  persistent analysis table directly instead of duplicating it into the
  per-query table session.
• Clause freshening (%freshen-clause, %freshen-term) and indexed-
  substitution's copy map use a small inline-array store that only escalates
  to a hash table past 12 entries, avoiding a hash-table allocation for the
  common case of a clause with few variables.
• Unification's cyclic-pair scratch space stores a nested eq hash table of
  "already seen" right-hand terms per left-hand term instead of a plain list
  checked with member, turning an O(n) scan per remembered pair into an
  O(1) lookup.
• Tabled answers now record whether they contain unbound variables; replaying
  a ground answer to a consumer skips variant instantiation entirely, since
  there is nothing to rename.
• Predicate first-argument indexing builds its per-key candidate lists with
  one linear merge over the original clause order instead of rescanning
  every clause once per distinct first-argument key, removing an
  O(clauses × distinct keys) reconstruction from %make-predicate-descriptor.
• Query-variable collection uses an eq hash-table membership check instead
  of pushnew, avoiding an O(n) rescan per occurrence, and the query-solving
  loop now collects a query's variables once before the search begins
  instead of once per solution.
• format/2's column directives (~t, ~|, ~+) track the pending
  segment's length and fill count incrementally instead of resumming the
  pending output on every column-fill request, and literal text runs batch
  through subseq instead of appending one character at a time.
• predsort/3's merge sort splits its private list destructively via
  nthcdr instead of allocating two subseq copies per recursion level; the
  caller copies the substituted input list once up front so the destructive
  split cannot alias shared structure.
• The lexer's tokenizer tracks open-delimiter depth with an explicit counter
  instead of recomputing it from the delimiter stack's length on every
  open/close paren, and its near-duplicate quoted-atom/string scanners are
  unified into one macro.
• paredit-cli bumped v0.8.0 -> v1.0.0.
• CI: the nix-setup composite action no longer receives CACHIX_AUTH_TOKEN
  on pull-request runs, so a fork's workflow changes cannot exfiltrate the
  write-capable Cachix credential; PR runs still read from the cache.
• CI: the release workflow's .asd-version and CHANGELOG.md-section
  extraction moved from sed/awk to Perl, which behaves identically on the
  GNU and BSD tool variants the supported runners carry (the prior sed
  pattern relied on a GNU-only escape and silently produced no match under
  macOS's BSD sed).

Fixed

• atom_number/2 and the number_string-family conversions parse the full
  ISO number-token grammar — including the 0x/0o/0b radix and 0'c
  character-code notations — by delegating to the same reader the tokenizer
  uses, instead of a separate hand-rolled decimal-only parser that rejected
  those notations.