odenwald, branch master

Merge pull request #1 from pks/claude

2026-02-27T11:09:49Z

Claude

Ignore output directories

2026-02-27T08:46:14Z

Co-Authored-By: Claude Opus 4.6

Fix C++ ow binary to produce translations

2026-02-26T18:31:35Z

- Node::mark and Node::score uninitialized, causing segfaults in topological_sort — add default initializers (0, 0.0) - odenwald.cc called incomplete sv_path() + exit(1) instead of viterbi_path() - viterbi_path: add reset() before topological_sort, initialize best_edge to nullptr - derive: off-by-one in NT order indexing — start j at 1 and use order[j]-1 (1-indexed alignment map) - read: ifs.readsome() returns 0 on macOS — use ifs.read() + ifs.gcount() - manual() signature: add missing Vocabulary parameter - Remove gperftools/tcmalloc dependency from Makefile Co-Authored-By: Claude Opus 4.6

Fix two bugs in prototype parser

2026-02-26T18:29:18Z

1. Inner visit at span (0,1) yielded no sub-spans because visit(1, 0, 1, 1) iterates span from 1 to r-x=0, which is empty. This prevented unary rules like [S] ||| [X,1] from completing at the leftmost span, so S(0,1) was never created. Drop the x=1 parameter (default x=0); scan already handles bounds checking. 2. Self-filling step searched remaining_items for unary NT rules, but those rules could be absent if consumed (advanced) during the parse loop. Look up grammar.start_nt directly instead, which covers all cases. Co-Authored-By: Claude Opus 4.6

Add Rust implementation of SCFG decoder

2026-02-26T18:28:22Z

Rust port of the Ruby prototype decoder with performance optimizations for real Hiero-style grammars: - Rule indexing by first terminal/NT symbol for fast lookup - Chart symbol interning (u16 IDs) instead of string hashing - Passive chart index by (symbol, left) for direct right-endpoint lookup - Items store rule index instead of cloned rule data Includes CKY+ parser, chart-to-hypergraph conversion, Viterbi decoding, derivation extraction, and JSON hypergraph I/O. Self-filling step in parse uses grammar lookup (not just remaining active items) to handle rules that were consumed during the parse loop or skipped by the has_any_at optimization. Produces identical output to the Ruby prototype on all test examples. Co-Authored-By: Claude Opus 4.6

Replace silent rescue with explicit type check in Item constructor

2026-02-24T16:18:29Z

When creating an Item from a Rule (not an Item), tail_spans doesn't exist. Check with is_a?(Item) instead of catching the exception silently. Co-Authored-By: Claude Opus 4.6

Emit binary glue rule only once

2026-02-24T16:16:06Z

The [S] -> [S] [X] concatenation rule was duplicated for every non-S LHS symbol. Move it out of the loop so it's added once. Co-Authored-By: Claude Opus 4.6

Deduplicate all_paths by reachable edge set

2026-02-24T16:14:37Z

The Cartesian product over all nodes produces duplicate derivations when edges differ only in nodes unreachable from the top. Walk reachable edges from the top edge of each path and drop paths with identical reachable sets. Co-Authored-By: Claude Opus 4.6

Fix reordering bug in derive and add test example

2026-02-24T16:07:57Z

derive used a sequential counter to index into the source-side NT map, which only worked for monotone rules. Now looks up tails by the target NT's own index via map.index(i.index). Adds toy-reorder example (German verb-final -> English SVO) to exercise the fix. Also updates trollop -> optimist and guards xmlsimple require. Co-Authored-By: Claude Opus 4.6

external/README

2017-08-04T19:24:05Z