WAM-CL ANSI test adapter

This directory adapts the maintained Paul Dietz ANSI Common Lisp tests in [`../ansi-test`](../ansi-test) to WAM-CL without modifying the vendored suite. The exact upstream repository, revision, and license are recorded in [upstream-revision.txt](upstream-revision.txt).

Current baseline

The first executable slice is ansi-test/iteration/loop.lsp. Its seven ordinary LOOP tests pass:

SLOOP.1 PASS
SLOOP.2 PASS
SLOOP.3 PASS
SLOOP.4 PASS
SLOOP.5 PASS
SLOOP.6 PASS
SLOOP.7 PASS

Run the slice from the repository root:

swipl -s t\ansi-test-wamcl\run.pl -g run_tests,halt -t "halt(1)"

The command fails if the child WAM-CL process fails, a selected test is missing, a test reports an error or wrong value, or not all seven result markers are present. It also enforces a three-minute child timeout and rejects any failed bootstrap form.

Each run writes a full Markdown transcript and a machine-readable summary to results/latest.md and results/latest.json. These generated files are ignored by Git; BASELINE.md records the reviewed baseline.

Adapter design

run.pl concatenates the required upstream bootstrap files and feeds them to a fresh WAM-CL process one form at a time. This avoids WAM-CL's current crash when these files are loaded through CL:LOAD.

child.pl evaluates registered tests only after the REPL reaches end of file. Dynamic EVAL currently interferes with an active stdin reader, so this keeps test execution isolated from source ingestion. It captures all returned values and compares them with the regression framework's EQUALP-WITH-CASE.

loop-prelude.lsp contains only the DEF-MACRO-TEST registration helper needed by iteration/loop.lsp. The generated error tests are intentionally not selected yet because they require SIGNALS-ERROR and complete condition handling. They must not be counted as passes until those facilities work.

Random auxiliary bootstrap

swipl -s t\ansi-test-wamcl\run-random.pl -g run_random_tests,halt -t "halt(1)"

This gate concatenates the unmodified shared sources through universe.lsp and auxiliary/random-aux.lsp, followed by random-bootstrap.lsp. It requires all 8 source-end markers, zero failed forms, and 9 passing assertions. The assertions exercise both macro expansion and execution of RCASE and RANDOM-CASE, cumulative weights, sequence selection, and universe values. The child timeout is ten minutes; results/random-latest.log is ignored.

The runtime fixes include checked random-state copying and isolation, INTEGER-LENGTH, BUTLAST, EVERY/SOME, and a bounded COERCE subset. ASSERT supports empty/omitted places, custom conditions, and retry through the CONTINUE restart; interactive place correction is not implemented and nonempty places are rejected. COERCE supports the simple string/list/vector type names, real-to-float conversions, character designators, and functions; this is not a complete implementation of compound type specifiers. Coercion preserves existing sequence identity when the supported target type already matches; lambda coercion creates a runtime closure that receives argument data without re-evaluating it. EVERY and SOME return exactly one value, even when their predicates return multiple values.

Compiler diagnostics no longer install printed, numbered-variable clauses. Top-level macros (including those in top-level PROGN, active EVAL-WHEN, and macro expansions) are installed by compilation, independent of verbosity. Nested DEFMACRO forms install their definitions only when executed. LIST allocates fresh cons cells at runtime instead of retaining compiler argument templates. TYPEP uses the correct argument order and subtype checks cannot backtrack through failed class lookups to spuriously succeed.

Next compliance barrier

Passing these gates does not establish full ANSI compliance or a successful load of the complete shared ansi-aux.lsp/symbol-name/notes bootstrap. The full auxiliary gate must be checked separately before enabling the maintained suite's generated macro-error tests.

A quiet, bounded full-load probe reached auxiliary/ansi-aux.lsp:548, `(defparameter displaced (make-int-array 100000))`, after seven complete upstream source markers and zero failed forms, but did not finish within 240 seconds. A prefix probe verified the base-, standard-, and code-character sequence lengths as 95, 96, and 256, respectively. Accessing those global strings with ELT still fails: globals can contain native Prolog strings, but f_elt/3 delegates to get_adata/2, which lacks that representation. The same probe also stalled on a three-element MAKE-INT-ARRAY; the full-load stall therefore cannot be attributed to large-array performance alone. Direct three-element construction, FUNCALL construction, and f_aset/4 storage probes succeed, leaving the Lisp-level array-update/iteration path to investigate. Neither this prefix nor the full shared bootstrap is a passing gate. Diagnostic transcripts are retained in ignored results/ansi-aux-latest.log and results/ansi-aux-prefix.log.

The project sanity suite currently passes 256/256 assertions (including 123 bootstrap assertions, up from 47); the maintained LOOP adapter remains 7/7, independently of the nine random-bootstrap assertions above.