Mine offline phrase library from rescue/dev-2026-07 (Tier 0)

MASTER_PLAN.md

@@ -371,3 +371,8 @@ on ours. License check before every download.
   debt (measured):** 16 transformer signs the avatar cannot play — BOOK, BUY, FEEL,
   HE, LEARN, LIKE, LIVE, MAKE, MILK, MY, NAME, PHONE, SOUTH AFRICA, TEACH, THINK,
   YOUR. Record real signer data for these (plus CALL, HERE) before expanding further.
+- 2026-07-05 — First dev-tag mining complete: offline phrase library (60 phrases,
+  Tier 0 exact-match before all other tiers) + contraction/SA-slang normalisation
+  ported from `rescue/dev-2026-07` into the rules-first pipeline. Copula 'AM'
+  glosses removed from the phrase data (copulas are not signed); the full file is
+  flagged for SASL interpreter review (D15). 8 new tests.

backend/data/offline_phrases.json

@@ -0,0 +1,63 @@
+{
+  "_comment": "Offline phrase → SASL gloss map, mined from tag rescue/dev-2026-07 (FEAT-2). English copula 'AM' glosses removed (copulas are not signed); all other sequences preserved verbatim. review_status: draft — entire file awaits SASL interpreter review (MASTER_PLAN D15).",
+  "where does it hurt": ["WHERE", "HURT", "YOU"],
+  "i need a doctor": ["DOCTOR", "NEED", "I"],
+  "call an ambulance": ["AMBULANCE", "CALL", "PLEASE"],
+  "i am in pain": ["PAIN", "I", "HAVE"],
+  "i cannot breathe": ["BREATHE", "CAN", "NOT", "I"],
+  "i am allergic": ["ALLERGIC", "I"],
+  "i am diabetic": ["DIABETIC", "I"],
+  "i take medication": ["MEDICINE", "I", "TAKE"],
+  "i am pregnant": ["PREGNANT", "I"],
+  "i feel dizzy": ["DIZZY", "I", "FEEL"],
+  "i feel nauseous": ["SICK", "I", "FEEL"],
+  "i have a fever": ["FEVER", "I", "HAVE"],
+  "my chest hurts": ["CHEST", "HURT", "MY"],
+  "i cannot walk": ["WALK", "CAN", "NOT", "I"],
+  "i need water": ["WATER", "NEED", "I"],
+  "i need food": ["FOOD", "NEED", "I"],
+  "i need the toilet": ["TOILET", "NEED", "I"],
+  "i need help": ["HELP", "NEED", "I"],
+  "please help me": ["HELP", "PLEASE", "ME"],
+  "call the police": ["POLICE", "CALL", "PLEASE"],
+  "i am deaf": ["DEAF", "I"],
+  "i use sign language": ["SIGN", "I", "USE"],
+  "please write it down": ["WRITE", "PLEASE"],
+  "please speak slowly": ["SLOW", "PLEASE", "SPEAK"],
+  "i do not understand": ["UNDERSTAND", "NOT", "I"],
+  "can you repeat that": ["REPEAT", "CAN", "YOU"],
+  "yes": ["YES"],
+  "no": ["NO"],
+  "thank you": ["THANK YOU"],
+  "please": ["PLEASE"],
+  "i agree": ["AGREE", "I"],
+  "i disagree": ["DISAGREE", "I"],
+  "i am scared": ["SCARED", "I"],
+  "i am angry": ["ANGRY", "I"],
+  "i am sad": ["SAD", "I"],
+  "i am happy": ["HAPPY", "I"],
+  "i am tired": ["TIRED", "I"],
+  "i am hungry": ["HUNGRY", "I"],
+  "i am thirsty": ["THIRSTY", "I"],
+  "i need my medication": ["MEDICINE", "MY", "NEED", "I"],
+  "i have a headache": ["HEAD", "HURT", "I", "HAVE"],
+  "my blood pressure is high": ["BLOOD", "PRESSURE", "HIGH"],
+  "i have asthma": ["ASTHMA", "I", "HAVE"],
+  "i have epilepsy": ["EPILEPSY", "I", "HAVE"],
+  "i am having a seizure": ["SEIZURE", "I", "HAVE"],
+  "i need an interpreter": ["INTERPRETER", "NEED", "I"],
+  "this is an emergency": ["EMERGENCY"],
+  "i am lost": ["LOST", "I"],
+  "where is the hospital": ["HOSPITAL", "WHERE"],
+  "where is the bathroom": ["TOILET", "WHERE"],
+  "how much does it cost": ["COST", "HOW MUCH"],
+  "i cannot afford it": ["MONEY", "NOT", "ENOUGH", "I"],
+  "i need a wheelchair": ["WHEELCHAIR", "NEED", "I"],
+  "i have rights": ["RIGHTS", "I", "HAVE"],
+  "i want a lawyer": ["LAWYER", "WANT", "I"],
+  "i do not consent": ["CONSENT", "NOT", "I"],
+  "stop": ["STOP"],
+  "wait": ["WAIT"],
+  "come here": ["COME", "HERE"],
+  "go away": ["GO", "AWAY"]
+}

backend/services/sasl_pipeline.py

@@ -19,9 +19,11 @@
 All AI runs locally via Ollama — no cloud API keys needed.
 """
 
+import json
 import logging
 import os
 import time as _time
+from pathlib import Path
 
 logger = logging.getLogger(__name__)
 
@@ -35,6 +37,68 @@
 # rules leave sign coverage below this fraction of tokens.
 LLM_ASSIST_COVERAGE_THRESHOLD = 0.70
 
+# ── Offline phrase library (mined from tag rescue/dev-2026-07, FEAT-2) ──────
+# Exact-match phrase → SASL gloss sequence, checked before any other tier.
+# Loaded once at import time from backend/data/offline_phrases.json.
+_OFFLINE_PHRASES: dict[str, list[str]] = {}
+_OFFLINE_PHRASES_PATH = Path(__file__).resolve().parent.parent / "data" / "offline_phrases.json"
+
+
+def _load_offline_phrases() -> None:
+    global _OFFLINE_PHRASES
+    try:
+        with open(_OFFLINE_PHRASES_PATH, encoding="utf-8") as fh:
+            data = json.load(fh)
+        _OFFLINE_PHRASES = {k: v for k, v in data.items() if not k.startswith("_")}
+        logger.info("[SASL] Loaded %d offline phrases", len(_OFFLINE_PHRASES))
+    except Exception as exc:
+        logger.warning("[SASL] Could not load offline phrases: %s", exc)
+
+
+_load_offline_phrases()
+
+# ── Contraction / SA slang normalisation map (mined from rescue/dev-2026-07) ─
+_CONTRACTIONS: dict[str, str] = {
+    "i'm": "i am", "i'll": "i will", "i've": "i have", "i'd": "i would",
+    "you're": "you are", "you'll": "you will", "you've": "you have",
+    "he's": "he is", "she's": "she is", "it's": "it is",
+    "we're": "we are", "we'll": "we will", "we've": "we have",
+    "they're": "they are", "they'll": "they will", "they've": "they have",
+    "don't": "do not", "doesn't": "does not", "didn't": "did not",
+    "won't": "will not", "wouldn't": "would not", "couldn't": "could not",
+    "shouldn't": "should not", "can't": "cannot", "isn't": "is not",
+    "aren't": "are not", "wasn't": "was not", "weren't": "were not",
+    "haven't": "have not", "hasn't": "has not", "hadn't": "had not",
+    "gonna": "going to", "wanna": "want to", "gotta": "got to",
+    # SA informal words
+    "howzit": "hello how are you",
+    "eish": "", "yoh": "",
+    "lekker": "good", "sharp": "okay",
+    "ja": "yes", "nee": "no",
+}
+
+
+def _normalize_informal(text: str) -> str:
+    """Expand contractions and normalise SA informal words before the pipeline.
+
+    Multi-word keys are replaced as substrings first (longest first); then
+    single-word tokens are expanded. Empty expansions drop the word.
+    """
+    text_lower = text.lower()
+    for phrase, expansion in sorted(_CONTRACTIONS.items(), key=lambda x: len(x[0]), reverse=True):
+        if " " in phrase and phrase in text_lower:
+            text_lower = text_lower.replace(phrase, expansion if expansion else "", 1)
+
+    words = text_lower.split()
+    result = []
+    for w in words:
+        expanded = _CONTRACTIONS.get(w)
+        if expanded is None:
+            result.append(w)
+        elif expanded:
+            result.extend(expanded.split())
+    return " ".join(result)
+
 # ── FEAT-5: Multilingual constants ─────────────────────────────────────────
 
 # Ollama base URL and model for translation (reuse shared env vars)
@@ -208,6 +272,24 @@ def _build_result(signs, gloss_text, english, **extras):
     # Import once — used by both tier 1 and tier 2
     from sasl_transformer.models import TranslationRequest
 
+    # ── Step 0: normalise informal input, then exact offline phrase match ─
+    # Checked after normalisation so contractions expand first:
+    # "I'm in pain!" → "i am in pain" → matches the offline entry.
+    text = _normalize_informal(text)
+    _phrase_key = text.lower().strip(" .!?,")
+    if _phrase_key in _OFFLINE_PHRASES:
+        _signs = _OFFLINE_PHRASES[_phrase_key]
+        _gloss = " ".join(_signs)
+        _lib = _sasl_transformer.sign_library
+        _known = sum(1 for s in _signs if _lib.has_sign(s))
+        _coverage = round(_known / len(_signs), 3) if _signs else 1.0
+        logger.info("[SASL] Tier 0 (phrase) '%s' → '%s' coverage=%.2f", _phrase_key, _gloss, _coverage)
+        return _build_result(
+            _signs, _gloss, text,
+            sign_coverage=_coverage,
+            fingerspelled=[s for s in _signs if not _lib.has_sign(s)],
+        )
+
     # ── D4 (MASTER_PLAN): deterministic rules run FIRST. ──────────────────
     # The LLM is an optional assist consulted only when rules leave too many
     # words unsigned, and its output is used only if it measurably improves

tests/test_offline_phrases.py

@@ -0,0 +1,66 @@
+"""
+Tests for the offline phrase library tier (mined from rescue/dev-2026-07).
+
+Proves:
+- exact phrase match bypasses all other tiers (Tier 0)
+- contraction/SA-slang normalisation feeds the phrase match
+- punctuation does not break matching
+- no copula glosses (AM/IS/ARE) remain in the phrase data
+- non-phrase input still flows to the rules tier
+
+Run with: pytest tests/test_offline_phrases.py -v
+No Ollama needed — phrase and rules tiers are fully offline.
+"""
+
+import asyncio
+
+from backend.services import sasl_pipeline
+
+
+def _run(text):
+    return asyncio.run(sasl_pipeline.text_to_sasl_signs(text))
+
+
+def test_phrases_loaded():
+    assert len(sasl_pipeline._OFFLINE_PHRASES) >= 50
+
+
+def test_exact_phrase_match():
+    result = _run("I need help")
+    assert result["signs"] == ["HELP", "NEED", "I"]
+
+
+def test_punctuation_does_not_break_match():
+    result = _run("I need help!")
+    assert result["signs"] == ["HELP", "NEED", "I"]
+
+
+def test_contraction_normalisation_feeds_phrase_match():
+    """'I'm hungry' must expand to 'i am hungry' and hit the phrase entry."""
+    result = _run("I'm hungry")
+    assert result["signs"] == ["HUNGRY", "I"]
+
+
+def test_sa_slang_normalisation():
+    """'Howzit' expands to the greeting phrase before translation."""
+    result = _run("Howzit")
+    assert "HELLO" in result["signs"]
+
+
+def test_no_copula_glosses_in_phrase_data():
+    """English copulas are not SASL signs and must never be glossed."""
+    for phrase, signs in sasl_pipeline._OFFLINE_PHRASES.items():
+        assert "AM" not in signs, f"{phrase!r} glosses the copula AM"
+        assert "IS" not in signs, f"{phrase!r} glosses the copula IS"
+        assert "ARE" not in signs, f"{phrase!r} glosses the copula ARE"
+
+
+def test_non_phrase_falls_through_to_rules():
+    result = _run("Tomorrow I will go to the doctor")
+    assert result["signs"][0] == "TOMORROW", "rules tier should still apply time-first order"
+    assert "WILL" in result["signs"]
+
+
+def test_emergency_phrase():
+    result = _run("This is an emergency")
+    assert result["signs"] == ["EMERGENCY"]