Compile cli_args.pl through the JS WAM (A2)

[s243a]

Compile the frozen A1 peerhailer argparser (examples/cli_args/cli_args.pl) with the wam_javascript interpreter target and hold it to the same oracle bar as A1.

What

• Import A1 examples/cli_args/ (oracle, harness, patternjs) from claude/peerhailer-exploratory-docs-aodas5. Top-level reference files are verbatim.
• examples/cli_args/wamjs/: build.sh regenerates js/ from cli_args.pl; ESM shim cliArgs.mjs exports parseArgs / CliError (term conversion only); corpus copy cliArgs.wamjs.test.mjs; diff_runner_wamjs.mjs + run_differential_wamjs.sh.

Runtime gaps this program forced

1. sub_string/5 — starts_with/2 and substring helpers call it; no user label, so the interpreter must implement the builtin (V.String Sub). Probe: probe_sub_string/0.
2. Y-save across Call — default_registry/1 is a huge ground fact (GetList/Unify*, no Allocate) and was clobbering the caller’s Y (including Result). Call snapshots Y; Proceed restores. Probe: probe_y_preserve/0.
3. Execute of a builtin must Proceed, not halt — helpers such as substring_from/3 end with Deallocate+Execute sub_string/5. Halting dropped the rest of parse_args/3. Probe: probe_tail_builtin/0.

The shim implements no parse rule. /3 custom registries are converted at the edge; the 17-test corpus only uses /2 + the default registry.

Acceptance (measured)

• bash examples/cli_args/wamjs/build.sh regenerates js/
• node --test examples/cli_args/wamjs/cliArgs.wamjs.test.mjs → 17/17
• Differential vs oracle, same seed:

sample size:          5067 argv-lines
  oracle ok results:  4150
  oracle errors:      917
divergences:          0
message mismatches:   0
timing: oracle 0.048s  wamjs 4.639s

• tests/test_wam_javascript_builtins.pl (incl. the three new probes), lowered, fact-sources, and CONFORMANCE_TARGETS=javascript → green

Does not edit core/target_registry.pl, the conformance harness, wam_target.pl, or frozen examples/cli_args/ top-level files.