fix: seven ISC pipeline conformance bugs — the corpus sweep runs end-to-end

[ronaldtse]

Closes the ISC corpus sweep enablement and lands the first tranche of conformance fixes it exposed.

The leak that failed iso-mal

Items::Set.from_strings(arr.map(&:to_s)) stringified every any([...]) list entry. A primitive entry — boundary inside any([boundary, "‌", "‍"]) — stringified to its object inspect, and the composed constraint became a garbage character class, so iso-mal-Mlym-Latn-15919-2001's boundary-anchored rules never fired. Lists now keep their Items; the adapter unwraps each entry kind. iso-mal passes 11/11 through the Interpreter and Compiler::Ruby with the stock ASCII \b boundary — the TS-boundary-divergence theory from the earlier analysis was a red herring caused by measuring against the wrong corpus.

The sweep was measuring the wrong corpus

Interscript.maps() enumerated the installed gem's frozen 2.4.x .imp corpus, not the maps checkout. The ISC_SWEEP gate now enumerates through the load path (matching locate()) and honors INTERSCRIPT_MAPS_PATH; before that, the "ISC sweep" ran 287 legacy .imp maps and the "six failures" were that corpus's iso-mal, unfixable from this repo.

Three conformance bugs the real sweep exposed

1. Ranges exploded into String#succ sequences. Expanding Items::Range to an array of single chars iterated Ruby's string succ ladder (a, b, … z, aa, ab …), which never reaches non-ASCII codepoints — any("a".."�") matched nothing beyond ASCII. This silently disabled every corpus map's word-capitalization post-rule (sub any("a".."�"), upcase, before: boundary): 73% of the 391 baseline failures were case-only (īş vs Īş). Ranges now stay native Any(Range); all three runtimes compile that to the codepoint class [lo-hi].
2. Escape sequences silently vanished. The Normalizer folded :string captures itself and read only the :char key, so \", \\, \n, \t and \uXXXX fragments decoded to nothing (bas-rus "pod"ezd", gost's hard-sign-to-quote rules). The Transform already had the correct fold; both paths now share Transform.decode_string_parts.
3. Every transliterate() call re-parsed the map. Two bypasses — the .isc branch of DSL.parse and Compiler.call's direct parse_isc — re-ran the full Parslet parse per call. A 700 KB map costs ~30 s per parse (the legacy .imp path parses the same map in 0.17 s), which blew the 5 s per-example test timeout. Both entry points now route through the one cached parse.
4. Bare alias refs inside any(...) leaked a debug string. any(greek) parsed as an :alias_ref no transform rule matched, degrading to the literal string [:alias_ref, "greek"] inside constraint regexps — every context-gated rule (Greek velar assimilation γ→n, ta-marbuta contexts) silently never fired. Alias refs now transform, and single-alias any() args keep set semantics matching the legacy runtime's Any(nil) for imported aliases (resolving them to the charset string would bake a 1000-char literal into a lookbehind that never matches).
5. ISC documents weren't addressable like .imp maps. Two parity gaps, both fatal to Compiler::Ruby: Document#name now derives from the file name (registering under the system code made every map whose code differs from its file name resolve to a Hash-default empty entry — stages[:main].call crashed; this was 3106 of the sweep's failures), and every alias/stage now carries doc_name, which the compilers emit as Maps.transliterate(stage.doc_name, ...) — without it, aliased-dependency runs (un-ell runs elot, 131 maps) compiled to transliterate(nil, s, :main).

Sweep status: opt-in, not a CI gate

The sweep runs end-to-end on all 289 ISC maps. It still measures a conformance tail, so ISC_SWEEP=1 stays env-gated for local tracking:

ISC_SWEEP=1 INTERSCRIPT_MAPS_PATH=../maps/maps \
  rspec spec/interscript_spec.rb

Flip it into the rake workflow once the tail reaches zero. The non-sweep suite is green with CI-equivalent flags.