feat: quality gate pipeline validation (#623)

Validates JSONL/JSON pipeline outputs for:
- Schema correctness
- Content quality (non-empty, not duplicated)
- Toxicity detection
- Dedup hash management with auto-cleanup

Usage:
  python3 bin/quality-gate.py validate data.jsonl
  python3 bin/quality-gate.py score data.jsonl
  python3 bin/quality-gate.py stats
  python3 bin/quality-gate.py cleanup

Closes #623

.hermes/training-data/scene-descriptions-jazz.jsonl

@@ -1,100 +0,0 @@
-{"song": "Blue in Green \u2014 Miles Davis", "beat": 1, "lyric_line": "Blue in green, a color no one sees", "scene": {"mood": "melancholy", "colors": ["navy", "silver"], "composition": "low angle", "camera": "slow pan", "description": "[Blue in Green, beat 1] Blue in green, a color no one sees \u2014 low angle shot, slow pan movement, palette shifts toward navy."}}
-{"song": "Blue in Green \u2014 Miles Davis", "beat": 2, "lyric_line": "The smoke curls upward like a question", "scene": {"mood": "melancholy", "colors": ["silver", "steel blue"], "composition": "high angle", "camera": "slow zoom", "description": "[Blue in Green, beat 2] The smoke curls upward like a question \u2014 high angle shot, slow zoom movement, palette shifts toward silver."}}
-{"song": "Blue in Green \u2014 Miles Davis", "beat": 3, "lyric_line": "The piano drops notes into silence", "scene": {"mood": "melancholy", "colors": ["steel blue", "slate gray"], "composition": "center frame", "camera": "dolly in", "description": "[Blue in Green, beat 3] The piano drops notes into silence \u2014 center frame shot, dolly in movement, palette shifts toward steel blue."}}
-{"song": "Blue in Green \u2014 Miles Davis", "beat": 4, "lyric_line": "Where does the melody go when it ends?", "scene": {"mood": "melancholy", "colors": ["slate gray", "navy"], "composition": "rule of thirds", "camera": "static wide", "description": "[Blue in Green, beat 4] Where does the melody go when it ends? \u2014 rule of thirds shot, static wide movement, palette shifts toward slate gray."}}
-{"song": "Blue in Green \u2014 Miles Davis", "beat": 5, "lyric_line": "A woman at the bar stares into her glass", "scene": {"mood": "melancholy", "colors": ["navy", "silver"], "composition": "extreme close-up", "camera": "handheld drift", "description": "[Blue in Green, beat 5] A woman at the bar stares into her glass \u2014 extreme close-up shot, handheld drift movement, palette shifts toward navy."}}
-{"song": "Blue in Green \u2014 Miles Davis", "beat": 6, "lyric_line": "The trumpet speaks without words", "scene": {"mood": "melancholy", "colors": ["silver", "steel blue"], "composition": "wide establishing", "camera": "crane down", "description": "[Blue in Green, beat 6] The trumpet speaks without words \u2014 wide establishing shot, crane down movement, palette shifts toward silver."}}
-{"song": "Blue in Green \u2014 Miles Davis", "beat": 7, "lyric_line": "Rain on the window, fingers on keys", "scene": {"mood": "melancholy", "colors": ["steel blue", "slate gray"], "composition": "over-the-shoulder", "camera": "tracking shot", "description": "[Blue in Green, beat 7] Rain on the window, fingers on keys \u2014 over-the-shoulder shot, tracking shot movement, palette shifts toward steel blue."}}
-{"song": "Blue in Green \u2014 Miles Davis", "beat": 8, "lyric_line": "Something beautiful is leaving", "scene": {"mood": "melancholy", "colors": ["slate gray", "navy"], "composition": "symmetrical", "camera": "close-up hold", "description": "[Blue in Green, beat 8] Something beautiful is leaving \u2014 symmetrical shot, close-up hold movement, palette shifts toward slate gray."}}
-{"song": "Blue in Green \u2014 Miles Davis", "beat": 9, "lyric_line": "The bass walks alone through the dark", "scene": {"mood": "melancholy", "colors": ["navy", "silver"], "composition": "dutch angle", "camera": "pull back", "description": "[Blue in Green, beat 9] The bass walks alone through the dark \u2014 dutch angle shot, pull back movement, palette shifts toward navy."}}
-{"song": "Blue in Green \u2014 Miles Davis", "beat": 10, "lyric_line": "Morning comes but the blue remains", "scene": {"mood": "melancholy", "colors": ["silver", "steel blue"], "composition": "silhouette", "camera": "orbit", "description": "[Blue in Green, beat 10] Morning comes but the blue remains \u2014 silhouette shot, orbit movement, palette shifts toward silver."}}
-{"song": "Strange Fruit \u2014 Billie Holiday", "beat": 1, "lyric_line": "Southern trees bear a strange fruit", "scene": {"mood": "protest", "colors": ["crimson", "black"], "composition": "low angle", "camera": "slow pan", "description": "[Strange Fruit, beat 1] Southern trees bear a strange fruit \u2014 low angle shot, slow pan movement, palette shifts toward crimson."}}
-{"song": "Strange Fruit \u2014 Billie Holiday", "beat": 2, "lyric_line": "Blood on the leaves and blood at the root", "scene": {"mood": "protest", "colors": ["black", "burnt orange"], "composition": "high angle", "camera": "slow zoom", "description": "[Strange Fruit, beat 2] Blood on the leaves and blood at the root \u2014 high angle shot, slow zoom movement, palette shifts toward black."}}
-{"song": "Strange Fruit \u2014 Billie Holiday", "beat": 3, "lyric_line": "Black bodies swinging in the southern breeze", "scene": {"mood": "protest", "colors": ["burnt orange", "blood red"], "composition": "center frame", "camera": "dolly in", "description": "[Strange Fruit, beat 3] Black bodies swinging in the southern breeze \u2014 center frame shot, dolly in movement, palette shifts toward burnt orange."}}
-{"song": "Strange Fruit \u2014 Billie Holiday", "beat": 4, "lyric_line": "Strange fruit hanging from the poplar trees", "scene": {"mood": "protest", "colors": ["blood red", "crimson"], "composition": "rule of thirds", "camera": "static wide", "description": "[Strange Fruit, beat 4] Strange fruit hanging from the poplar trees \u2014 rule of thirds shot, static wide movement, palette shifts toward blood red."}}
-{"song": "Strange Fruit \u2014 Billie Holiday", "beat": 5, "lyric_line": "Pastoral scene of the gallant south", "scene": {"mood": "protest", "colors": ["crimson", "black"], "composition": "extreme close-up", "camera": "handheld drift", "description": "[Strange Fruit, beat 5] Pastoral scene of the gallant south \u2014 extreme close-up shot, handheld drift movement, palette shifts toward crimson."}}
-{"song": "Strange Fruit \u2014 Billie Holiday", "beat": 6, "lyric_line": "The bulging eyes and the twisted mouth", "scene": {"mood": "protest", "colors": ["black", "burnt orange"], "composition": "wide establishing", "camera": "crane down", "description": "[Strange Fruit, beat 6] The bulging eyes and the twisted mouth \u2014 wide establishing shot, crane down movement, palette shifts toward black."}}
-{"song": "Strange Fruit \u2014 Billie Holiday", "beat": 7, "lyric_line": "Scent of magnolia sweet and fresh", "scene": {"mood": "protest", "colors": ["burnt orange", "blood red"], "composition": "over-the-shoulder", "camera": "tracking shot", "description": "[Strange Fruit, beat 7] Scent of magnolia sweet and fresh \u2014 over-the-shoulder shot, tracking shot movement, palette shifts toward burnt orange."}}
-{"song": "Strange Fruit \u2014 Billie Holiday", "beat": 8, "lyric_line": "Then the sudden smell of burning flesh", "scene": {"mood": "protest", "colors": ["blood red", "crimson"], "composition": "symmetrical", "camera": "close-up hold", "description": "[Strange Fruit, beat 8] Then the sudden smell of burning flesh \u2014 symmetrical shot, close-up hold movement, palette shifts toward blood red."}}
-{"song": "Strange Fruit \u2014 Billie Holiday", "beat": 9, "lyric_line": "Here is a fruit for the crows to pluck", "scene": {"mood": "protest", "colors": ["crimson", "black"], "composition": "dutch angle", "camera": "pull back", "description": "[Strange Fruit, beat 9] Here is a fruit for the crows to pluck \u2014 dutch angle shot, pull back movement, palette shifts toward crimson."}}
-{"song": "Strange Fruit \u2014 Billie Holiday", "beat": 10, "lyric_line": "For the rain to gather, for the wind to suck", "scene": {"mood": "protest", "colors": ["black", "burnt orange"], "composition": "silhouette", "camera": "orbit", "description": "[Strange Fruit, beat 10] For the rain to gather, for the wind to suck \u2014 silhouette shot, orbit movement, palette shifts toward black."}}
-{"song": "Take Five \u2014 Dave Brubeck", "beat": 1, "lyric_line": "Take five, the saxophone says", "scene": {"mood": "cool", "colors": ["slate blue", "charcoal"], "composition": "low angle", "camera": "slow pan", "description": "[Take Five, beat 1] Take five, the saxophone says \u2014 low angle shot, slow pan movement, palette shifts toward slate blue."}}
-{"song": "Take Five \u2014 Dave Brubeck", "beat": 2, "lyric_line": "Fingers move in patterns only jazz knows", "scene": {"mood": "cool", "colors": ["charcoal", "ivory"], "composition": "high angle", "camera": "slow zoom", "description": "[Take Five, beat 2] Fingers move in patterns only jazz knows \u2014 high angle shot, slow zoom movement, palette shifts toward charcoal."}}
-{"song": "Take Five \u2014 Dave Brubeck", "beat": 3, "lyric_line": "The rhythm counts in fives and sevens", "scene": {"mood": "cool", "colors": ["ivory", "copper"], "composition": "center frame", "camera": "dolly in", "description": "[Take Five, beat 3] The rhythm counts in fives and sevens \u2014 center frame shot, dolly in movement, palette shifts toward ivory."}}
-{"song": "Take Five \u2014 Dave Brubeck", "beat": 4, "lyric_line": "Coffee steam rises with the bass line", "scene": {"mood": "cool", "colors": ["copper", "slate blue"], "composition": "rule of thirds", "camera": "static wide", "description": "[Take Five, beat 4] Coffee steam rises with the bass line \u2014 rule of thirds shot, static wide movement, palette shifts toward copper."}}
-{"song": "Take Five \u2014 Dave Brubeck", "beat": 5, "lyric_line": "A newspaper folds in half, unread", "scene": {"mood": "cool", "colors": ["slate blue", "charcoal"], "composition": "extreme close-up", "camera": "handheld drift", "description": "[Take Five, beat 5] A newspaper folds in half, unread \u2014 extreme close-up shot, handheld drift movement, palette shifts toward slate blue."}}
-{"song": "Take Five \u2014 Dave Brubeck", "beat": 6, "lyric_line": "The piano answers what the sax asked", "scene": {"mood": "cool", "colors": ["charcoal", "ivory"], "composition": "wide establishing", "camera": "crane down", "description": "[Take Five, beat 6] The piano answers what the sax asked \u2014 wide establishing shot, crane down movement, palette shifts toward charcoal."}}
-{"song": "Take Five \u2014 Dave Brubeck", "beat": 7, "lyric_line": "Time stretches like an afternoon", "scene": {"mood": "cool", "colors": ["ivory", "copper"], "composition": "over-the-shoulder", "camera": "tracking shot", "description": "[Take Five, beat 7] Time stretches like an afternoon \u2014 over-the-shoulder shot, tracking shot movement, palette shifts toward ivory."}}
-{"song": "Take Five \u2014 Dave Brubeck", "beat": 8, "lyric_line": "Nobody here is in a hurry", "scene": {"mood": "cool", "colors": ["copper", "slate blue"], "composition": "symmetrical", "camera": "close-up hold", "description": "[Take Five, beat 8] Nobody here is in a hurry \u2014 symmetrical shot, close-up hold movement, palette shifts toward copper."}}
-{"song": "Take Five \u2014 Dave Brubeck", "beat": 9, "lyric_line": "The drummer smiles at nobody", "scene": {"mood": "cool", "colors": ["slate blue", "charcoal"], "composition": "dutch angle", "camera": "pull back", "description": "[Take Five, beat 9] The drummer smiles at nobody \u2014 dutch angle shot, pull back movement, palette shifts toward slate blue."}}
-{"song": "Take Five \u2014 Dave Brubeck", "beat": 10, "lyric_line": "Five beats to the measure, forever", "scene": {"mood": "cool", "colors": ["charcoal", "ivory"], "composition": "silhouette", "camera": "orbit", "description": "[Take Five, beat 10] Five beats to the measure, forever \u2014 silhouette shot, orbit movement, palette shifts toward charcoal."}}
-{"song": "Round Midnight \u2014 Thelonious Monk", "beat": 1, "lyric_line": "It begins to tell round midnight", "scene": {"mood": "solitude", "colors": ["midnight blue", "amber"], "composition": "low angle", "camera": "slow pan", "description": "[Round Midnight, beat 1] It begins to tell round midnight \u2014 low angle shot, slow pan movement, palette shifts toward midnight blue."}}
-{"song": "Round Midnight \u2014 Thelonious Monk", "beat": 2, "lyric_line": "The city exhales its last commuters", "scene": {"mood": "solitude", "colors": ["amber", "smoke gray"], "composition": "high angle", "camera": "slow zoom", "description": "[Round Midnight, beat 2] The city exhales its last commuters \u2014 high angle shot, slow zoom movement, palette shifts toward amber."}}
-{"song": "Round Midnight \u2014 Thelonious Monk", "beat": 3, "lyric_line": "A single light in a window above", "scene": {"mood": "solitude", "colors": ["smoke gray", "deep purple"], "composition": "center frame", "camera": "dolly in", "description": "[Round Midnight, beat 3] A single light in a window above \u2014 center frame shot, dolly in movement, palette shifts toward smoke gray."}}
-{"song": "Round Midnight \u2014 Thelonious Monk", "beat": 4, "lyric_line": "The piano waits for the right touch", "scene": {"mood": "solitude", "colors": ["deep purple", "midnight blue"], "composition": "rule of thirds", "camera": "static wide", "description": "[Round Midnight, beat 4] The piano waits for the right touch \u2014 rule of thirds shot, static wide movement, palette shifts toward deep purple."}}
-{"song": "Round Midnight \u2014 Thelonious Monk", "beat": 5, "lyric_line": "Dissonance is just another kind of truth", "scene": {"mood": "solitude", "colors": ["midnight blue", "amber"], "composition": "extreme close-up", "camera": "handheld drift", "description": "[Round Midnight, beat 5] Dissonance is just another kind of truth \u2014 extreme close-up shot, handheld drift movement, palette shifts toward midnight blue."}}
-{"song": "Round Midnight \u2014 Thelonious Monk", "beat": 6, "lyric_line": "The shadows play jazz on the walls", "scene": {"mood": "solitude", "colors": ["amber", "smoke gray"], "composition": "wide establishing", "camera": "crane down", "description": "[Round Midnight, beat 6] The shadows play jazz on the walls \u2014 wide establishing shot, crane down movement, palette shifts toward amber."}}
-{"song": "Round Midnight \u2014 Thelonious Monk", "beat": 7, "lyric_line": "A cigarette burns down to the filter", "scene": {"mood": "solitude", "colors": ["smoke gray", "deep purple"], "composition": "over-the-shoulder", "camera": "tracking shot", "description": "[Round Midnight, beat 7] A cigarette burns down to the filter \u2014 over-the-shoulder shot, tracking shot movement, palette shifts toward smoke gray."}}
-{"song": "Round Midnight \u2014 Thelonious Monk", "beat": 8, "lyric_line": "Nobody calls after midnight", "scene": {"mood": "solitude", "colors": ["deep purple", "midnight blue"], "composition": "symmetrical", "camera": "close-up hold", "description": "[Round Midnight, beat 8] Nobody calls after midnight \u2014 symmetrical shot, close-up hold movement, palette shifts toward deep purple."}}
-{"song": "Round Midnight \u2014 Thelonious Monk", "beat": 9, "lyric_line": "The melody walks crooked through the dark", "scene": {"mood": "solitude", "colors": ["midnight blue", "amber"], "composition": "dutch angle", "camera": "pull back", "description": "[Round Midnight, beat 9] The melody walks crooked through the dark \u2014 dutch angle shot, pull back movement, palette shifts toward midnight blue."}}
-{"song": "Round Midnight \u2014 Thelonious Monk", "beat": 10, "lyric_line": "Round midnight, the world is honest", "scene": {"mood": "solitude", "colors": ["amber", "smoke gray"], "composition": "silhouette", "camera": "orbit", "description": "[Round Midnight, beat 10] Round midnight, the world is honest \u2014 silhouette shot, orbit movement, palette shifts toward amber."}}
-{"song": "Summertime \u2014 Ella Fitzgerald", "beat": 1, "lyric_line": "Summertime and the livin is easy", "scene": {"mood": "yearning", "colors": ["golden", "warm amber"], "composition": "low angle", "camera": "slow pan", "description": "[Summertime, beat 1] Summertime and the livin is easy \u2014 low angle shot, slow pan movement, palette shifts toward golden."}}
-{"song": "Summertime \u2014 Ella Fitzgerald", "beat": 2, "lyric_line": "Fish are jumpin and the cotton is high", "scene": {"mood": "yearning", "colors": ["warm amber", "pale blue"], "composition": "high angle", "camera": "slow zoom", "description": "[Summertime, beat 2] Fish are jumpin and the cotton is high \u2014 high angle shot, slow zoom movement, palette shifts toward warm amber."}}
-{"song": "Summertime \u2014 Ella Fitzgerald", "beat": 3, "lyric_line": "Oh your daddy is rich and your ma is good lookin", "scene": {"mood": "yearning", "colors": ["pale blue", "cream"], "composition": "center frame", "camera": "dolly in", "description": "[Summertime, beat 3] Oh your daddy is rich and your ma is good lookin \u2014 center frame shot, dolly in movement, palette shifts toward pale blue."}}
-{"song": "Summertime \u2014 Ella Fitzgerald", "beat": 4, "lyric_line": "So hush little baby dont you cry", "scene": {"mood": "yearning", "colors": ["cream", "golden"], "composition": "rule of thirds", "camera": "static wide", "description": "[Summertime, beat 4] So hush little baby dont you cry \u2014 rule of thirds shot, static wide movement, palette shifts toward cream."}}
-{"song": "Summertime \u2014 Ella Fitzgerald", "beat": 5, "lyric_line": "One of these mornings you gonna rise up singing", "scene": {"mood": "yearning", "colors": ["golden", "warm amber"], "composition": "extreme close-up", "camera": "handheld drift", "description": "[Summertime, beat 5] One of these mornings you gonna rise up singing \u2014 extreme close-up shot, handheld drift movement, palette shifts toward golden."}}
-{"song": "Summertime \u2014 Ella Fitzgerald", "beat": 6, "lyric_line": "Then you spread your wings and you take to the sky", "scene": {"mood": "yearning", "colors": ["warm amber", "pale blue"], "composition": "wide establishing", "camera": "crane down", "description": "[Summertime, beat 6] Then you spread your wings and you take to the sky \u2014 wide establishing shot, crane down movement, palette shifts toward warm amber."}}
-{"song": "Summertime \u2014 Ella Fitzgerald", "beat": 7, "lyric_line": "But till that morning theres nothin can harm you", "scene": {"mood": "yearning", "colors": ["pale blue", "cream"], "composition": "over-the-shoulder", "camera": "tracking shot", "description": "[Summertime, beat 7] But till that morning theres nothin can harm you \u2014 over-the-shoulder shot, tracking shot movement, palette shifts toward pale blue."}}
-{"song": "Summertime \u2014 Ella Fitzgerald", "beat": 8, "lyric_line": "With daddy and mammy standin by", "scene": {"mood": "yearning", "colors": ["cream", "golden"], "composition": "symmetrical", "camera": "close-up hold", "description": "[Summertime, beat 8] With daddy and mammy standin by \u2014 symmetrical shot, close-up hold movement, palette shifts toward cream."}}
-{"song": "Summertime \u2014 Ella Fitzgerald", "beat": 9, "lyric_line": "The river moves slow in the delta heat", "scene": {"mood": "yearning", "colors": ["golden", "warm amber"], "composition": "dutch angle", "camera": "pull back", "description": "[Summertime, beat 9] The river moves slow in the delta heat \u2014 dutch angle shot, pull back movement, palette shifts toward golden."}}
-{"song": "Summertime \u2014 Ella Fitzgerald", "beat": 10, "lyric_line": "A screen door bangs somewhere in the distance", "scene": {"mood": "yearning", "colors": ["warm amber", "pale blue"], "composition": "silhouette", "camera": "orbit", "description": "[Summertime, beat 10] A screen door bangs somewhere in the distance \u2014 silhouette shot, orbit movement, palette shifts toward warm amber."}}
-{"song": "A Love Supreme \u2014 John Coltrane", "beat": 1, "lyric_line": "A love supreme, the saxophone prays", "scene": {"mood": "spiritual", "colors": ["white", "gold"], "composition": "low angle", "camera": "slow pan", "description": "[A Love Supreme, beat 1] A love supreme, the saxophone prays \u2014 low angle shot, slow pan movement, palette shifts toward white."}}
-{"song": "A Love Supreme \u2014 John Coltrane", "beat": 2, "lyric_line": "Notes rise like incense through the smoke", "scene": {"mood": "spiritual", "colors": ["gold", "deep violet"], "composition": "high angle", "camera": "slow zoom", "description": "[A Love Supreme, beat 2] Notes rise like incense through the smoke \u2014 high angle shot, slow zoom movement, palette shifts toward gold."}}
-{"song": "A Love Supreme \u2014 John Coltrane", "beat": 3, "lyric_line": "The drummer is a congregation", "scene": {"mood": "spiritual", "colors": ["deep violet", "burning orange"], "composition": "center frame", "camera": "dolly in", "description": "[A Love Supreme, beat 3] The drummer is a congregation \u2014 center frame shot, dolly in movement, palette shifts toward deep violet."}}
-{"song": "A Love Supreme \u2014 John Coltrane", "beat": 4, "lyric_line": "God lives in the space between notes", "scene": {"mood": "spiritual", "colors": ["burning orange", "white"], "composition": "rule of thirds", "camera": "static wide", "description": "[A Love Supreme, beat 4] God lives in the space between notes \u2014 rule of thirds shot, static wide movement, palette shifts toward burning orange."}}
-{"song": "A Love Supreme \u2014 John Coltrane", "beat": 5, "lyric_line": "The bass walks toward something holy", "scene": {"mood": "spiritual", "colors": ["white", "gold"], "composition": "extreme close-up", "camera": "handheld drift", "description": "[A Love Supreme, beat 5] The bass walks toward something holy \u2014 extreme close-up shot, handheld drift movement, palette shifts toward white."}}
-{"song": "A Love Supreme \u2014 John Coltrane", "beat": 6, "lyric_line": "Each phrase is a step closer to truth", "scene": {"mood": "spiritual", "colors": ["gold", "deep violet"], "composition": "wide establishing", "camera": "crane down", "description": "[A Love Supreme, beat 6] Each phrase is a step closer to truth \u2014 wide establishing shot, crane down movement, palette shifts toward gold."}}
-{"song": "A Love Supreme \u2014 John Coltrane", "beat": 7, "lyric_line": "The sax screams what words cannot hold", "scene": {"mood": "spiritual", "colors": ["deep violet", "burning orange"], "composition": "over-the-shoulder", "camera": "tracking shot", "description": "[A Love Supreme, beat 7] The sax screams what words cannot hold \u2014 over-the-shoulder shot, tracking shot movement, palette shifts toward deep violet."}}
-{"song": "A Love Supreme \u2014 John Coltrane", "beat": 8, "lyric_line": "There is no audience, only witness", "scene": {"mood": "spiritual", "colors": ["burning orange", "white"], "composition": "symmetrical", "camera": "close-up hold", "description": "[A Love Supreme, beat 8] There is no audience, only witness \u2014 symmetrical shot, close-up hold movement, palette shifts toward burning orange."}}
-{"song": "A Love Supreme \u2014 John Coltrane", "beat": 9, "lyric_line": "The music burns like a sermon on fire", "scene": {"mood": "spiritual", "colors": ["white", "gold"], "composition": "dutch angle", "camera": "pull back", "description": "[A Love Supreme, beat 9] The music burns like a sermon on fire \u2014 dutch angle shot, pull back movement, palette shifts toward white."}}
-{"song": "A Love Supreme \u2014 John Coltrane", "beat": 10, "lyric_line": "A love supreme, a love supreme", "scene": {"mood": "spiritual", "colors": ["gold", "deep violet"], "composition": "silhouette", "camera": "orbit", "description": "[A Love Supreme, beat 10] A love supreme, a love supreme \u2014 silhouette shot, orbit movement, palette shifts toward gold."}}
-{"song": "Feeling Good \u2014 Nina Simone", "beat": 1, "lyric_line": "Birds flying high you know how I feel", "scene": {"mood": "liberation", "colors": ["bright yellow", "sky blue"], "composition": "low angle", "camera": "slow pan", "description": "[Feeling Good, beat 1] Birds flying high you know how I feel \u2014 low angle shot, slow pan movement, palette shifts toward bright yellow."}}
-{"song": "Feeling Good \u2014 Nina Simone", "beat": 2, "lyric_line": "Sun in the sky you know how I feel", "scene": {"mood": "liberation", "colors": ["sky blue", "emerald green"], "composition": "high angle", "camera": "slow zoom", "description": "[Feeling Good, beat 2] Sun in the sky you know how I feel \u2014 high angle shot, slow zoom movement, palette shifts toward sky blue."}}
-{"song": "Feeling Good \u2014 Nina Simone", "beat": 3, "lyric_line": "Breeze driftin on by you know how I feel", "scene": {"mood": "liberation", "colors": ["emerald green", "sunrise orange"], "composition": "center frame", "camera": "dolly in", "description": "[Feeling Good, beat 3] Breeze driftin on by you know how I feel \u2014 center frame shot, dolly in movement, palette shifts toward emerald green."}}
-{"song": "Feeling Good \u2014 Nina Simone", "beat": 4, "lyric_line": "Its a new dawn its a new day its a new life", "scene": {"mood": "liberation", "colors": ["sunrise orange", "bright yellow"], "composition": "rule of thirds", "camera": "static wide", "description": "[Feeling Good, beat 4] Its a new dawn its a new day its a new life \u2014 rule of thirds shot, static wide movement, palette shifts toward sunrise orange."}}
-{"song": "Feeling Good \u2014 Nina Simone", "beat": 5, "lyric_line": "And Im feeling good", "scene": {"mood": "liberation", "colors": ["bright yellow", "sky blue"], "composition": "extreme close-up", "camera": "handheld drift", "description": "[Feeling Good, beat 5] And Im feeling good \u2014 extreme close-up shot, handheld drift movement, palette shifts toward bright yellow."}}
-{"song": "Feeling Good \u2014 Nina Simone", "beat": 6, "lyric_line": "Fish in the sea you know how I feel", "scene": {"mood": "liberation", "colors": ["sky blue", "emerald green"], "composition": "wide establishing", "camera": "crane down", "description": "[Feeling Good, beat 6] Fish in the sea you know how I feel \u2014 wide establishing shot, crane down movement, palette shifts toward sky blue."}}
-{"song": "Feeling Good \u2014 Nina Simone", "beat": 7, "lyric_line": "River running free you know how I feel", "scene": {"mood": "liberation", "colors": ["emerald green", "sunrise orange"], "composition": "over-the-shoulder", "camera": "tracking shot", "description": "[Feeling Good, beat 7] River running free you know how I feel \u2014 over-the-shoulder shot, tracking shot movement, palette shifts toward emerald green."}}
-{"song": "Feeling Good \u2014 Nina Simone", "beat": 8, "lyric_line": "Blossom on the tree you know how I feel", "scene": {"mood": "liberation", "colors": ["sunrise orange", "bright yellow"], "composition": "symmetrical", "camera": "close-up hold", "description": "[Feeling Good, beat 8] Blossom on the tree you know how I feel \u2014 symmetrical shot, close-up hold movement, palette shifts toward sunrise orange."}}
-{"song": "Feeling Good \u2014 Nina Simone", "beat": 9, "lyric_line": "Dragonfly out in the sun you know what I mean", "scene": {"mood": "liberation", "colors": ["bright yellow", "sky blue"], "composition": "dutch angle", "camera": "pull back", "description": "[Feeling Good, beat 9] Dragonfly out in the sun you know what I mean \u2014 dutch angle shot, pull back movement, palette shifts toward bright yellow."}}
-{"song": "Feeling Good \u2014 Nina Simone", "beat": 10, "lyric_line": "Butterflies all havin fun you know what I mean", "scene": {"mood": "liberation", "colors": ["sky blue", "emerald green"], "composition": "silhouette", "camera": "orbit", "description": "[Feeling Good, beat 10] Butterflies all havin fun you know what I mean \u2014 silhouette shot, orbit movement, palette shifts toward sky blue."}}
-{"song": "My Funny Valentine \u2014 Chet Baker", "beat": 1, "lyric_line": "My funny valentine, sweet comic valentine", "scene": {"mood": "tender", "colors": ["rose", "soft pink"], "composition": "low angle", "camera": "slow pan", "description": "[My Funny Valentine, beat 1] My funny valentine, sweet comic valentine \u2014 low angle shot, slow pan movement, palette shifts toward rose."}}
-{"song": "My Funny Valentine \u2014 Chet Baker", "beat": 2, "lyric_line": "You make me smile with my heart", "scene": {"mood": "tender", "colors": ["soft pink", "warm cream"], "composition": "high angle", "camera": "slow zoom", "description": "[My Funny Valentine, beat 2] You make me smile with my heart \u2014 high angle shot, slow zoom movement, palette shifts toward soft pink."}}
-{"song": "My Funny Valentine \u2014 Chet Baker", "beat": 3, "lyric_line": "Your looks are laughable unphotographable", "scene": {"mood": "tender", "colors": ["warm cream", "dusty lavender"], "composition": "center frame", "camera": "dolly in", "description": "[My Funny Valentine, beat 3] Your looks are laughable unphotographable \u2014 center frame shot, dolly in movement, palette shifts toward warm cream."}}
-{"song": "My Funny Valentine \u2014 Chet Baker", "beat": 4, "lyric_line": "Yet youre my favorite work of art", "scene": {"mood": "tender", "colors": ["dusty lavender", "rose"], "composition": "rule of thirds", "camera": "static wide", "description": "[My Funny Valentine, beat 4] Yet youre my favorite work of art \u2014 rule of thirds shot, static wide movement, palette shifts toward dusty lavender."}}
-{"song": "My Funny Valentine \u2014 Chet Baker", "beat": 5, "lyric_line": "Is your figure less than greek", "scene": {"mood": "tender", "colors": ["rose", "soft pink"], "composition": "extreme close-up", "camera": "handheld drift", "description": "[My Funny Valentine, beat 5] Is your figure less than greek \u2014 extreme close-up shot, handheld drift movement, palette shifts toward rose."}}
-{"song": "My Funny Valentine \u2014 Chet Baker", "beat": 6, "lyric_line": "Is your mouth a little weak", "scene": {"mood": "tender", "colors": ["soft pink", "warm cream"], "composition": "wide establishing", "camera": "crane down", "description": "[My Funny Valentine, beat 6] Is your mouth a little weak \u2014 wide establishing shot, crane down movement, palette shifts toward soft pink."}}
-{"song": "My Funny Valentine \u2014 Chet Baker", "beat": 7, "lyric_line": "When you open it to speak are you smart", "scene": {"mood": "tender", "colors": ["warm cream", "dusty lavender"], "composition": "over-the-shoulder", "camera": "tracking shot", "description": "[My Funny Valentine, beat 7] When you open it to speak are you smart \u2014 over-the-shoulder shot, tracking shot movement, palette shifts toward warm cream."}}
-{"song": "My Funny Valentine \u2014 Chet Baker", "beat": 8, "lyric_line": "Dont change a hair for me", "scene": {"mood": "tender", "colors": ["dusty lavender", "rose"], "composition": "symmetrical", "camera": "close-up hold", "description": "[My Funny Valentine, beat 8] Dont change a hair for me \u2014 symmetrical shot, close-up hold movement, palette shifts toward dusty lavender."}}
-{"song": "My Funny Valentine \u2014 Chet Baker", "beat": 9, "lyric_line": "Not if you care for me", "scene": {"mood": "tender", "colors": ["rose", "soft pink"], "composition": "dutch angle", "camera": "pull back", "description": "[My Funny Valentine, beat 9] Not if you care for me \u2014 dutch angle shot, pull back movement, palette shifts toward rose."}}
-{"song": "My Funny Valentine \u2014 Chet Baker", "beat": 10, "lyric_line": "Stay little valentine stay", "scene": {"mood": "tender", "colors": ["soft pink", "warm cream"], "composition": "silhouette", "camera": "orbit", "description": "[My Funny Valentine, beat 10] Stay little valentine stay \u2014 silhouette shot, orbit movement, palette shifts toward soft pink."}}
-{"song": "So What \u2014 Miles Davis", "beat": 1, "lyric_line": "So what, the trumpet shrugs", "scene": {"mood": "detached", "colors": ["cool gray", "steel"], "composition": "low angle", "camera": "slow pan", "description": "[So What, beat 1] So what, the trumpet shrugs \u2014 low angle shot, slow pan movement, palette shifts toward cool gray."}}
-{"song": "So What \u2014 Miles Davis", "beat": 2, "lyric_line": "Two chords is all you need", "scene": {"mood": "detached", "colors": ["steel", "pale blue"], "composition": "high angle", "camera": "slow zoom", "description": "[So What, beat 2] Two chords is all you need \u2014 high angle shot, slow zoom movement, palette shifts toward steel."}}
-{"song": "So What \u2014 Miles Davis", "beat": 3, "lyric_line": "The bass walks like it knows where its going", "scene": {"mood": "detached", "colors": ["pale blue", "frost white"], "composition": "center frame", "camera": "dolly in", "description": "[So What, beat 3] The bass walks like it knows where its going \u2014 center frame shot, dolly in movement, palette shifts toward pale blue."}}
-{"song": "So What \u2014 Miles Davis", "beat": 4, "lyric_line": "But the destination is irrelevant", "scene": {"mood": "detached", "colors": ["frost white", "cool gray"], "composition": "rule of thirds", "camera": "static wide", "description": "[So What, beat 4] But the destination is irrelevant \u2014 rule of thirds shot, static wide movement, palette shifts toward frost white."}}
-{"song": "So What \u2014 Miles Davis", "beat": 5, "lyric_line": "The piano comps and stays out of the way", "scene": {"mood": "detached", "colors": ["cool gray", "steel"], "composition": "extreme close-up", "camera": "handheld drift", "description": "[So What, beat 5] The piano comps and stays out of the way \u2014 extreme close-up shot, handheld drift movement, palette shifts toward cool gray."}}
-{"song": "So What \u2014 Miles Davis", "beat": 6, "lyric_line": "Cool means never raising your voice", "scene": {"mood": "detached", "colors": ["steel", "pale blue"], "composition": "wide establishing", "camera": "crane down", "description": "[So What, beat 6] Cool means never raising your voice \u2014 wide establishing shot, crane down movement, palette shifts toward steel."}}
-{"song": "So What \u2014 Miles Davis", "beat": 7, "lyric_line": "The sax takes a breath and keeps walking", "scene": {"mood": "detached", "colors": ["pale blue", "frost white"], "composition": "over-the-shoulder", "camera": "tracking shot", "description": "[So What, beat 7] The sax takes a breath and keeps walking \u2014 over-the-shoulder shot, tracking shot movement, palette shifts toward pale blue."}}
-{"song": "So What \u2014 Miles Davis", "beat": 8, "lyric_line": "Nothing is urgent here", "scene": {"mood": "detached", "colors": ["frost white", "cool gray"], "composition": "symmetrical", "camera": "close-up hold", "description": "[So What, beat 8] Nothing is urgent here \u2014 symmetrical shot, close-up hold movement, palette shifts toward frost white."}}
-{"song": "So What \u2014 Miles Davis", "beat": 9, "lyric_line": "The smoke moves slower than the music", "scene": {"mood": "detached", "colors": ["cool gray", "steel"], "composition": "dutch angle", "camera": "pull back", "description": "[So What, beat 9] The smoke moves slower than the music \u2014 dutch angle shot, pull back movement, palette shifts toward cool gray."}}
-{"song": "So What \u2014 Miles Davis", "beat": 10, "lyric_line": "So what. So what indeed.", "scene": {"mood": "detached", "colors": ["steel", "pale blue"], "composition": "silhouette", "camera": "orbit", "description": "[So What, beat 10] So what. So what indeed. \u2014 silhouette shot, orbit movement, palette shifts toward steel."}}
-{"song": "Gloomy Sunday \u2014 Billie Holiday", "beat": 1, "lyric_line": "Gloomy Sunday with shadows I spend it all", "scene": {"mood": "despair", "colors": ["black", "deep crimson"], "composition": "low angle", "camera": "slow pan", "description": "[Gloomy Sunday, beat 1] Gloomy Sunday with shadows I spend it all \u2014 low angle shot, slow pan movement, palette shifts toward black."}}
-{"song": "Gloomy Sunday \u2014 Billie Holiday", "beat": 2, "lyric_line": "My heart and I have decided to end it all", "scene": {"mood": "despair", "colors": ["deep crimson", "midnight"], "composition": "high angle", "camera": "slow zoom", "description": "[Gloomy Sunday, beat 2] My heart and I have decided to end it all \u2014 high angle shot, slow zoom movement, palette shifts toward deep crimson."}}
-{"song": "Gloomy Sunday \u2014 Billie Holiday", "beat": 3, "lyric_line": "Soon therell be candles and prayers that are sad I know", "scene": {"mood": "despair", "colors": ["midnight", "ash gray"], "composition": "center frame", "camera": "dolly in", "description": "[Gloomy Sunday, beat 3] Soon therell be candles and prayers that are sad I know \u2014 center frame shot, dolly in movement, palette shifts toward midnight."}}
-{"song": "Gloomy Sunday \u2014 Billie Holiday", "beat": 4, "lyric_line": "Let them not weep let them know that Im glad to go", "scene": {"mood": "despair", "colors": ["ash gray", "black"], "composition": "rule of thirds", "camera": "static wide", "description": "[Gloomy Sunday, beat 4] Let them not weep let them know that Im glad to go \u2014 rule of thirds shot, static wide movement, palette shifts toward ash gray."}}
-{"song": "Gloomy Sunday \u2014 Billie Holiday", "beat": 5, "lyric_line": "Death is no dream for in death Im caressing you", "scene": {"mood": "despair", "colors": ["black", "deep crimson"], "composition": "extreme close-up", "camera": "handheld drift", "description": "[Gloomy Sunday, beat 5] Death is no dream for in death Im caressing you \u2014 extreme close-up shot, handheld drift movement, palette shifts toward black."}}
-{"song": "Gloomy Sunday \u2014 Billie Holiday", "beat": 6, "lyric_line": "With the last breath of my soul Ill be blessing you", "scene": {"mood": "despair", "colors": ["deep crimson", "midnight"], "composition": "wide establishing", "camera": "crane down", "description": "[Gloomy Sunday, beat 6] With the last breath of my soul Ill be blessing you \u2014 wide establishing shot, crane down movement, palette shifts toward deep crimson."}}
-{"song": "Gloomy Sunday \u2014 Billie Holiday", "beat": 7, "lyric_line": "Gloomy Sunday", "scene": {"mood": "despair", "colors": ["midnight", "ash gray"], "composition": "over-the-shoulder", "camera": "tracking shot", "description": "[Gloomy Sunday, beat 7] Gloomy Sunday \u2014 over-the-shoulder shot, tracking shot movement, palette shifts toward midnight."}}
-{"song": "Gloomy Sunday \u2014 Billie Holiday", "beat": 8, "lyric_line": "Dreaming I was only dreaming", "scene": {"mood": "despair", "colors": ["ash gray", "black"], "composition": "symmetrical", "camera": "close-up hold", "description": "[Gloomy Sunday, beat 8] Dreaming I was only dreaming \u2014 symmetrical shot, close-up hold movement, palette shifts toward ash gray."}}
-{"song": "Gloomy Sunday \u2014 Billie Holiday", "beat": 9, "lyric_line": "I wake and I find you asleep in the deep of my heart here", "scene": {"mood": "despair", "colors": ["black", "deep crimson"], "composition": "dutch angle", "camera": "pull back", "description": "[Gloomy Sunday, beat 9] I wake and I find you asleep in the deep of my heart here \u2014 dutch angle shot, pull back movement, palette shifts toward black."}}
-{"song": "Gloomy Sunday \u2014 Billie Holiday", "beat": 10, "lyric_line": "Darling I hope that my dream never haunted you", "scene": {"mood": "despair", "colors": ["deep crimson", "midnight"], "composition": "silhouette", "camera": "orbit", "description": "[Gloomy Sunday, beat 10] Darling I hope that my dream never haunted you \u2014 silhouette shot, orbit movement, palette shifts toward deep crimson."}}

bin/glitch_patterns.py

@@ -31,6 +31,14 @@ class GlitchCategory(Enum):
     WATER_REFLECTION = "water_reflection"
     SKYBOX_SEAM = "skybox_seam"
 
+    # Three.js-specific categories (ref: timmy-config#543)
+    SHADER_FAILURE = "shader_failure"
+    TEXTURE_PLACEHOLDER = "texture_placeholder"
+    UV_MAPPING_ERROR = "uv_mapping_error"
+    FRUSTUM_CULLING = "frustum_culling"
+    SHADOW_MAP_ARTIFACT = "shadow_map_artifact"
+    BLOOM_OVERFLOW = "bloom_overflow"
+
 
 @dataclass
 class GlitchPattern:
@@ -241,6 +249,123 @@ MATRIX_GLITCH_PATTERNS: list[GlitchPattern] = [
         ],
         confidence_threshold=0.45,
     ),
+
+    # --- Three.js-Specific Glitch Patterns (ref: timmy-config#543) ---
+    GlitchPattern(
+        category=GlitchCategory.SHADER_FAILURE,
+        name="Shader Compilation Failure",
+        description="Three.js shader failed to compile, rendering the material as solid black. "
+                    "Common when custom ShaderMaterial has syntax errors or missing uniforms.",
+        severity=GlitchSeverity.CRITICAL,
+        detection_prompts=[
+            "Look for objects or surfaces rendered as pure black (#000000) that should have visible textures or materials.",
+            "Identify geometry that appears completely dark while surrounding objects are normally lit.",
+            "Check for objects where the material seems to 'absorb all light' — flat black with no shading gradient.",
+        ],
+        visual_indicators=[
+            "solid black object with no shading",
+            "geometry rendered as silhouette",
+            "material appears to absorb light entirely",
+            "black patch inconsistent with scene lighting",
+        ],
+        confidence_threshold=0.7,
+    ),
+    GlitchPattern(
+        category=GlitchCategory.TEXTURE_PLACEHOLDER,
+        name="Three.js Texture Not Loaded",
+        description="Three.js failed to load the texture asset, rendering a 1x1 white pixel "
+                    "stretched across the entire surface. Distinguished from missing-texture by "
+                    "the uniform white/grey appearance rather than magenta.",
+        severity=GlitchSeverity.CRITICAL,
+        detection_prompts=[
+            "Look for surfaces that are uniformly white or light grey with no texture detail, even on large geometry.",
+            "Identify objects where the texture appears as a single solid color stretched across complex UVs.",
+            "Check for surfaces that look 'blank' or 'unloaded' — flat white/grey where detail should exist.",
+        ],
+        visual_indicators=[
+            "uniform white or light grey surface",
+            "no texture detail on large geometry",
+            "stretched single-color appearance",
+            "1x1 pixel placeholder stretched to fill UV space",
+        ],
+        confidence_threshold=0.65,
+    ),
+    GlitchPattern(
+        category=GlitchCategory.UV_MAPPING_ERROR,
+        name="BufferGeometry UV Mapping Error",
+        description="Three.js BufferGeometry has incorrect UV coordinates, causing textures to "
+                    "appear stretched, compressed, or mapped to the wrong faces.",
+        severity=GlitchSeverity.HIGH,
+        detection_prompts=[
+            "Look for textures that appear dramatically stretched in one direction on specific faces.",
+            "Identify surfaces where the texture pattern is distorted but other nearby surfaces look correct.",
+            "Check for faces where the texture seems 'smeared' or mapped with incorrect aspect ratio.",
+        ],
+        visual_indicators=[
+            "texture stretching on specific faces",
+            "distorted pattern on geometry",
+            "smeared texture appearance",
+            "aspect ratio mismatch between texture and surface",
+        ],
+        confidence_threshold=0.6,
+    ),
+    GlitchPattern(
+        category=GlitchCategory.FRUSTUM_CULLING,
+        name="Frustum Culling Artifact",
+        description="Three.js frustum culling incorrectly marks objects as outside the camera "
+                    "frustum, causing them to pop in/out of existence at screen edges.",
+        severity=GlitchSeverity.MEDIUM,
+        detection_prompts=[
+            "Look for objects that are partially visible at the edge of the frame — half-rendered or cut off unnaturally.",
+            "Identify geometry that seems to 'pop' into existence as the view angle changes.",
+            "Check screen edges for objects that appear suddenly rather than smoothly entering the viewport.",
+        ],
+        visual_indicators=[
+            "half-visible object at screen edge",
+            "object popping into frame",
+            "abrupt appearance of geometry",
+            "bounding box visible but mesh missing",
+        ],
+        confidence_threshold=0.55,
+    ),
+    GlitchPattern(
+        category=GlitchCategory.SHADOW_MAP_ARTIFACT,
+        name="Shadow Map Resolution Artifact",
+        description="Three.js shadow map has insufficient resolution, causing pixelated, "
+                    "blocky shadows with visible texel edges instead of smooth shadow gradients.",
+        severity=GlitchSeverity.MEDIUM,
+        detection_prompts=[
+            "Look for shadows with visible blocky or pixelated edges instead of smooth gradients.",
+            "Identify shadow maps where individual texels (texture pixels) are clearly visible.",
+            "Check for shadows that appear as jagged stair-stepped patterns rather than soft edges.",
+        ],
+        visual_indicators=[
+            "blocky shadow edges",
+            "visible texel grid in shadows",
+            "stair-stepped shadow boundary",
+            "pixelated shadow gradient",
+        ],
+        confidence_threshold=0.55,
+    ),
+    GlitchPattern(
+        category=GlitchCategory.BLOOM_OVERFLOW,
+        name="Post-Processing Bloom Overflow",
+        description="Three.js UnrealBloomPass or similar post-processing bloom effect is too "
+                    "intense, causing bright areas to bleed glow into surrounding geometry.",
+        severity=GlitchSeverity.LOW,
+        detection_prompts=[
+            "Look for bright areas that have an unusually large, soft glow bleeding into adjacent surfaces.",
+            "Identify scenes where light sources appear to have a 'halo' that extends beyond physical plausibility.",
+            "Check for bright objects whose glow color bleeds onto nearby unrelated geometry.",
+        ],
+        visual_indicators=[
+            "excessive glow bleeding from bright surfaces",
+            "halo around light sources",
+            "bloom color tinting adjacent geometry",
+            "glow bleeding beyond object boundaries",
+        ],
+        confidence_threshold=0.5,
+    ),
 ]
 
 
@@ -289,6 +414,23 @@ def build_vision_prompt(patterns: list[GlitchPattern] | None = None) -> str:
     )
 
 
+
+# Three.js-specific category set for filtering (ref: timmy-config#543)
+THREEJS_CATEGORIES = {
+    GlitchCategory.SHADER_FAILURE,
+    GlitchCategory.TEXTURE_PLACEHOLDER,
+    GlitchCategory.UV_MAPPING_ERROR,
+    GlitchCategory.FRUSTUM_CULLING,
+    GlitchCategory.SHADOW_MAP_ARTIFACT,
+    GlitchCategory.BLOOM_OVERFLOW,
+}
+
+
+def get_threejs_patterns() -> list[GlitchPattern]:
+    """Return only Three.js-specific glitch patterns."""
+    return [p for p in MATRIX_GLITCH_PATTERNS if p.category in THREEJS_CATEGORIES]
+
+
 if __name__ == "__main__":
     import json
     print(f"Loaded {len(MATRIX_GLITCH_PATTERNS)} glitch patterns:\n")

bin/matrix_glitch_detector.py

@@ -9,7 +9,7 @@ Usage:
     python matrix_glitch_detector.py <url> [--angles 4] [--output report.json]
     python matrix_glitch_detector.py --demo  # Run with synthetic test data
 
-Ref: timmy-config#491
+Ref: timmy-config#491, timmy-config#543
 """
 
 import argparse
@@ -33,6 +33,7 @@ from glitch_patterns import (
     MATRIX_GLITCH_PATTERNS,
     build_vision_prompt,
     get_patterns_by_severity,
+    get_threejs_patterns,
 )
 
 
@@ -345,14 +346,17 @@ def _parse_vision_response(
 
 def _infer_severity(category: str, confidence: float) -> str:
     """Infer severity from category and confidence when not provided."""
-    critical_cats = {"missing_textures", "clipping"}
-    high_cats = {"floating_assets", "broken_normals"}
+    critical_cats = {"missing_textures", "clipping", "shader_failure", "texture_placeholder"}
+    high_cats = {"floating_assets", "broken_normals", "uv_mapping_error"}
+    medium_cats = {"frustum_culling", "shadow_map_artifact"}
 
     cat_lower = category.lower()
     if any(c in cat_lower for c in critical_cats):
         return "critical" if confidence > 0.7 else "high"
     if any(c in cat_lower for c in high_cats):
         return "high" if confidence > 0.7 else "medium"
+    if any(c in cat_lower for c in medium_cats):
+        return "medium" if confidence > 0.6 else "low"
     return "medium" if confidence > 0.6 else "low"
 
 
@@ -389,9 +393,9 @@ def build_report(
             ),
         },
         metadata={
-            "detector_version": "0.1.0",
+            "detector_version": "0.2.0",
             "pattern_count": len(MATRIX_GLITCH_PATTERNS),
-            "reference": "timmy-config#491",
+            "reference": "timmy-config#491, timmy-config#543",
         },
     )
 
@@ -460,6 +464,30 @@ def run_demo(output_path: Optional[Path] = None) -> ScanResult:
             screenshot_index=3,
             screenshot_angle="left",
         ),
+        DetectedGlitch(
+            id=str(uuid.uuid4())[:8],
+            category="shader_failure",
+            name="Black Material on Portal Frame",
+            description="Portal frame rendered as solid black — shader compilation failed (missing uniform u_time)",
+            severity="critical",
+            confidence=0.91,
+            location_x=45.0,
+            location_y=30.0,
+            screenshot_index=0,
+            screenshot_angle="front",
+        ),
+        DetectedGlitch(
+            id=str(uuid.uuid4())[:8],
+            category="shadow_map_artifact",
+            name="Pixelated Character Shadow",
+            description="Character shadow shows visible texel grid — shadow map resolution too low (512x512)",
+            severity="medium",
+            confidence=0.78,
+            location_x=52.0,
+            location_y=75.0,
+            screenshot_index=1,
+            screenshot_angle="right",
+        ),
     ]
 
     print(f"[*] Detected {len(demo_glitches)} glitches")
@@ -496,6 +524,11 @@ Examples:
         help="Minimum severity to include in report",
     )
     parser.add_argument("--verbose", "-v", action="store_true", help="Verbose output")
+    parser.add_argument(
+        "--threejs",
+        action="store_true",
+        help="Focus on Three.js-specific glitch patterns only (shader, texture, UV, culling, shadow, bloom)",
+    )
 
     args = parser.parse_args()
 
@@ -525,9 +558,13 @@ Examples:
     screenshots = capture_screenshots(args.url, angles, screenshots_dir)
     print(f"[*] Captured {len(screenshots)} screenshots")
 
-    # Filter patterns by severity
+    # Filter patterns by severity and type
     min_sev = GlitchSeverity(args.min_severity)
     patterns = get_patterns_by_severity(min_sev)
+    if args.threejs:
+        threejs_patterns = get_threejs_patterns()
+        patterns = [p for p in patterns if p in threejs_patterns]
+        print(f"[*] Three.js-focused mode: {len(patterns)} patterns")
 
     # Analyze with vision AI
     print(f"[*] Analyzing with vision AI ({len(patterns)} patterns)...")

bin/nostr-agent-demo.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 """
 Full Nostr agent-to-agent communication demo - FINAL WORKING
 """

bin/preflight-provider-check.py

@@ -0,0 +1,271 @@
+#!/usr/bin/env python3
+"""
+Pre-Flight Provider Check Script
+Issue #508: [Robustness] Credential drain detection — provider health checks
+
+Pre-flight check before session launch: verifies provider credentials and balance.
+
+Usage:
+  python3 preflight-provider-check.py              # Check all providers
+  python3 preflight-provider-check.py --launch      # Check and return exit code
+  python3 preflight-provider-check.py --balance     # Check OpenRouter balance
+"""
+
+import os, sys, json, yaml, urllib.request
+from datetime import datetime, timezone
+from pathlib import Path
+
+# Configuration
+HERMES_HOME = Path(os.environ.get("HERMES_HOME", Path.home() / ".hermes"))
+LOG_DIR = Path.home() / ".local" / "timmy" / "fleet-health"
+LOG_FILE = LOG_DIR / "preflight-check.log"
+
+def log(msg):
+    """Log message to file and optionally console."""
+    timestamp = datetime.now(timezone.utc).strftime("%Y-%m-%d %H:%M:%S")
+    log_entry = "[" + timestamp + "] " + msg
+    
+    LOG_DIR.mkdir(parents=True, exist_ok=True)
+    with open(LOG_FILE, "a") as f:
+        f.write(log_entry + "\n")
+    
+    if "--quiet" not in sys.argv:
+        print(log_entry)
+
+def get_provider_api_key(provider):
+    """Get API key for a provider from .env or environment."""
+    env_file = HERMES_HOME / ".env"
+    if env_file.exists():
+        with open(env_file) as f:
+            for line in f:
+                line = line.strip()
+                if line.startswith(provider.upper() + "_API_KEY="):
+                    return line.split("=", 1)[1].strip().strip("'\"")
+    
+    return os.environ.get(provider.upper() + "_API_KEY")
+
+def check_openrouter_balance(api_key):
+    """Check OpenRouter balance via /api/v1/auth/key."""
+    if not api_key:
+        return False, "No API key", 0
+    
+    try:
+        req = urllib.request.Request(
+            "https://openrouter.ai/api/v1/auth/key",
+            headers={"Authorization": "Bearer " + api_key}
+        )
+        resp = urllib.request.urlopen(req, timeout=10)
+        data = json.loads(resp.read())
+        
+        # Check for credits
+        credits = data.get("data", {}).get("limit", 0)
+        usage = data.get("data", {}).get("usage", 0)
+        remaining = credits - usage if credits else None
+        
+        if remaining is not None and remaining <= 0:
+            return False, "No credits remaining", 0
+        elif remaining is not None:
+            return True, "Credits available", remaining
+        else:
+            return True, "Unlimited or unknown balance", None
+    
+    except urllib.error.HTTPError as e:
+        if e.code == 401:
+            return False, "Invalid API key", 0
+        else:
+            return False, "HTTP " + str(e.code), 0
+    except Exception as e:
+        return False, str(e)[:100], 0
+
+def check_nous_key(api_key):
+    """Check Nous API key with minimal test call."""
+    if not api_key:
+        return False, "No API key"
+    
+    try:
+        req = urllib.request.Request(
+            "https://inference.nousresearch.com/v1/models",
+            headers={"Authorization": "Bearer " + api_key}
+        )
+        resp = urllib.request.urlopen(req, timeout=10)
+        
+        if resp.status == 200:
+            return True, "Valid key"
+        else:
+            return False, "HTTP " + str(resp.status)
+    
+    except urllib.error.HTTPError as e:
+        if e.code == 401:
+            return False, "Invalid API key"
+        elif e.code == 403:
+            return False, "Forbidden"
+        else:
+            return False, "HTTP " + str(e.code)
+    except Exception as e:
+        return False, str(e)[:100]
+
+def check_anthropic_key(api_key):
+    """Check Anthropic API key with minimal test call."""
+    if not api_key:
+        return False, "No API key"
+    
+    try:
+        req = urllib.request.Request(
+            "https://api.anthropic.com/v1/models",
+            headers={
+                "x-api-key": api_key,
+                "anthropic-version": "2023-06-01"
+            }
+        )
+        resp = urllib.request.urlopen(req, timeout=10)
+        
+        if resp.status == 200:
+            return True, "Valid key"
+        else:
+            return False, "HTTP " + str(resp.status)
+    
+    except urllib.error.HTTPError as e:
+        if e.code == 401:
+            return False, "Invalid API key"
+        elif e.code == 403:
+            return False, "Forbidden"
+        else:
+            return False, "HTTP " + str(e.code)
+    except Exception as e:
+        return False, str(e)[:100]
+
+def check_ollama():
+    """Check if Ollama is running."""
+    try:
+        req = urllib.request.Request("http://localhost:11434/api/tags")
+        resp = urllib.request.urlopen(req, timeout=5)
+        
+        if resp.status == 200:
+            data = json.loads(resp.read())
+            models = data.get("models", [])
+            return True, str(len(models)) + " models loaded"
+        else:
+            return False, "HTTP " + str(resp.status)
+    
+    except Exception as e:
+        return False, str(e)[:100]
+
+def get_configured_provider():
+    """Get the configured provider from global config."""
+    config_file = HERMES_HOME / "config.yaml"
+    if not config_file.exists():
+        return None
+    
+    try:
+        with open(config_file) as f:
+            config = yaml.safe_load(f)
+        
+        model_config = config.get("model", {})
+        if isinstance(model_config, dict):
+            return model_config.get("provider")
+    except:
+        pass
+    
+    return None
+
+def run_preflight_check():
+    """Run pre-flight check on all providers."""
+    log("=== Pre-Flight Provider Check ===")
+    
+    results = {}
+    
+    # Check OpenRouter
+    or_key = get_provider_api_key("openrouter")
+    or_ok, or_msg, or_balance = check_openrouter_balance(or_key)
+    results["openrouter"] = {"healthy": or_ok, "message": or_msg, "balance": or_balance}
+    
+    # Check Nous
+    nous_key = get_provider_api_key("nous")
+    nous_ok, nous_msg = check_nous_key(nous_key)
+    results["nous"] = {"healthy": nous_ok, "message": nous_msg}
+    
+    # Check Anthropic
+    anthropic_key = get_provider_api_key("anthropic")
+    anthropic_ok, anthropic_msg = check_anthropic_key(anthropic_key)
+    results["anthropic"] = {"healthy": anthropic_ok, "message": anthropic_msg}
+    
+    # Check Ollama
+    ollama_ok, ollama_msg = check_ollama()
+    results["ollama"] = {"healthy": ollama_ok, "message": ollama_msg}
+    
+    # Get configured provider
+    configured = get_configured_provider()
+    
+    # Summary
+    healthy_count = sum(1 for r in results.values() if r["healthy"])
+    total_count = len(results)
+    
+    log("Results: " + str(healthy_count) + "/" + str(total_count) + " providers healthy")
+    
+    for provider, result in results.items():
+        status = "HEALTHY" if result["healthy"] else "UNHEALTHY"
+        extra = ""
+        if provider == "openrouter" and result.get("balance") is not None:
+            extra = " (balance: " + str(result["balance"]) + ")"
+        
+        log("  " + provider + ": " + status + " - " + result["message"] + extra)
+    
+    if configured:
+        log("Configured provider: " + configured)
+        if configured in results and not results[configured]["healthy"]:
+            log("WARNING: Configured provider " + configured + " is UNHEALTHY!")
+    
+    return results, configured
+
+def check_launch_readiness():
+    """Check if we're ready to launch sessions."""
+    results, configured = run_preflight_check()
+    
+    # Check if configured provider is healthy
+    if configured and configured in results:
+        if not results[configured]["healthy"]:
+            log("LAUNCH BLOCKED: Configured provider " + configured + " is unhealthy")
+            return False, configured + " is unhealthy"
+    
+    # Check if at least one provider is healthy
+    healthy_providers = [p for p, r in results.items() if r["healthy"]]
+    if not healthy_providers:
+        log("LAUNCH BLOCKED: No healthy providers available")
+        return False, "No healthy providers"
+    
+    log("LAUNCH READY: " + str(len(healthy_providers)) + " healthy providers available")
+    return True, "Ready"
+
+def show_balance():
+    """Show OpenRouter balance."""
+    api_key = get_provider_api_key("openrouter")
+    if not api_key:
+        print("No OpenRouter API key found")
+        return
+    
+    ok, msg, balance = check_openrouter_balance(api_key)
+    
+    if ok:
+        if balance is not None:
+            print("OpenRouter balance: " + str(balance) + " credits")
+        else:
+            print("OpenRouter: " + msg)
+    else:
+        print("OpenRouter: " + msg)
+
+def main():
+    if "--balance" in sys.argv:
+        show_balance()
+    elif "--launch" in sys.argv:
+        ready, message = check_launch_readiness()
+        if ready:
+            print("READY")
+            sys.exit(0)
+        else:
+            print("BLOCKED: " + message)
+            sys.exit(1)
+    else:
+        run_preflight_check()
+
+if __name__ == "__main__":
+    main()

bin/provider-health-monitor.py

@@ -0,0 +1,411 @@
+#!/usr/bin/env python3
+"""
+Provider Health Monitor Script
+Issue #509: [Robustness] Provider-aware profile config — auto-switch on failure
+
+Monitors provider health and automatically switches profiles to working providers.
+
+Usage:
+  python3 provider-health-monitor.py              # Run once
+  python3 provider-health-monitor.py --daemon     # Run continuously
+  python3 provider-health-monitor.py --status     # Show provider health
+"""
+
+import os, sys, json, yaml, urllib.request, time
+from datetime import datetime, timezone
+from pathlib import Path
+
+# Configuration
+HERMES_HOME = Path(os.environ.get("HERMES_HOME", Path.home() / ".hermes"))
+PROFILES_DIR = HERMES_HOME / "profiles"
+LOG_DIR = Path.home() / ".local" / "timmy" / "fleet-health"
+STATE_FILE = LOG_DIR / "tmux-state.json"
+LOG_FILE = LOG_DIR / "provider-health.log"
+
+# Provider test endpoints
+PROVIDER_TESTS = {
+    "openrouter": {
+        "url": "https://openrouter.ai/api/v1/models",
+        "method": "GET",
+        "headers": lambda api_key: {"Authorization": "Bearer " + api_key},
+        "timeout": 10
+    },
+    "anthropic": {
+        "url": "https://api.anthropic.com/v1/models",
+        "method": "GET",
+        "headers": lambda api_key: {"x-api-key": api_key, "anthropic-version": "2023-06-01"},
+        "timeout": 10
+    },
+    "nous": {
+        "url": "https://inference.nousresearch.com/v1/models",
+        "method": "GET",
+        "headers": lambda api_key: {"Authorization": "Bearer " + api_key},
+        "timeout": 10
+    },
+    "kimi-coding": {
+        "url": "https://api.kimi.com/coding/v1/models",
+        "method": "GET",
+        "headers": lambda api_key: {"x-api-key": api_key, "x-api-provider": "kimi-coding"},
+        "timeout": 10
+    },
+    "ollama": {
+        "url": "http://localhost:11434/api/tags",
+        "method": "GET",
+        "headers": lambda api_key: {},
+        "timeout": 5
+    }
+}
+
+def log(msg):
+    """Log message to file and optionally console."""
+    timestamp = datetime.now(timezone.utc).strftime("%Y-%m-%d %H:%M:%S")
+    log_entry = "[" + timestamp + "] " + msg
+    
+    LOG_DIR.mkdir(parents=True, exist_ok=True)
+    with open(LOG_FILE, "a") as f:
+        f.write(log_entry + "\n")
+    
+    if "--quiet" not in sys.argv:
+        print(log_entry)
+
+def get_provider_api_key(provider):
+    """Get API key for a provider from .env or environment."""
+    env_file = HERMES_HOME / ".env"
+    if env_file.exists():
+        with open(env_file) as f:
+            for line in f:
+                line = line.strip()
+                if line.startswith(provider.upper() + "_API_KEY="):
+                    return line.split("=", 1)[1].strip().strip("'\"")
+    
+    return os.environ.get(provider.upper() + "_API_KEY")
+
+def test_provider(provider, api_key=None):
+    """Test if a provider is healthy."""
+    config = PROVIDER_TESTS.get(provider)
+    if not config:
+        return False, "Unknown provider: " + provider
+    
+    headers = config["headers"](api_key or "")
+    
+    try:
+        req = urllib.request.Request(
+            config["url"],
+            headers=headers,
+            method=config["method"]
+        )
+        resp = urllib.request.urlopen(req, timeout=config["timeout"])
+        
+        if resp.status == 200:
+            return True, "Healthy"
+        else:
+            return False, "HTTP " + str(resp.status)
+    
+    except urllib.error.HTTPError as e:
+        if e.code == 401:
+            return False, "Unauthorized (401)"
+        elif e.code == 403:
+            return False, "Forbidden (403)"
+        elif e.code == 429:
+            return True, "Rate limited but accessible"
+        else:
+            return False, "HTTP " + str(e.code)
+    except Exception as e:
+        return False, str(e)[:100]
+
+def get_all_providers():
+    """Get all providers from profiles and global config."""
+    providers = set()
+    
+    # Global config
+    global_config = HERMES_HOME / "config.yaml"
+    if global_config.exists():
+        try:
+            with open(global_config) as f:
+                config = yaml.safe_load(f)
+            
+            # Primary model provider
+            model_config = config.get("model", {})
+            if isinstance(model_config, dict):
+                provider = model_config.get("provider", "")
+                if provider:
+                    providers.add(provider)
+            
+            # Auxiliary providers
+            auxiliary = config.get("auxiliary", {})
+            for aux_config in auxiliary.values():
+                if isinstance(aux_config, dict):
+                    provider = aux_config.get("provider", "")
+                    if provider and provider != "auto":
+                        providers.add(provider)
+        except:
+            pass
+    
+    # Profile configs
+    if PROFILES_DIR.exists():
+        for profile_dir in PROFILES_DIR.iterdir():
+            if profile_dir.is_dir():
+                config_file = profile_dir / "config.yaml"
+                if config_file.exists():
+                    try:
+                        with open(config_file) as f:
+                            config = yaml.safe_load(f)
+                        
+                        model_config = config.get("model", {})
+                        if isinstance(model_config, dict):
+                            provider = model_config.get("provider", "")
+                            if provider:
+                                providers.add(provider)
+                        
+                        auxiliary = config.get("auxiliary", {})
+                        for aux_config in auxiliary.values():
+                            if isinstance(aux_config, dict):
+                                provider = aux_config.get("provider", "")
+                                if provider and provider != "auto":
+                                    providers.add(provider)
+                    except:
+                        pass
+    
+    # Add common providers even if not configured
+    providers.update(["openrouter", "nous", "ollama"])
+    
+    return list(providers)
+
+def build_health_map():
+    """Build a health map of all providers."""
+    providers = get_all_providers()
+    health_map = {}
+    
+    log("Testing " + str(len(providers)) + " providers...")
+    
+    for provider in providers:
+        api_key = get_provider_api_key(provider)
+        healthy, message = test_provider(provider, api_key)
+        
+        health_map[provider] = {
+            "healthy": healthy,
+            "message": message,
+            "last_test": datetime.now(timezone.utc).isoformat(),
+            "api_key_present": bool(api_key)
+        }
+        
+        status = "HEALTHY" if healthy else "UNHEALTHY"
+        log("  " + provider + ": " + status + " - " + message)
+    
+    return health_map
+
+def get_fallback_providers(health_map):
+    """Get list of healthy providers in priority order."""
+    # Priority order: nous, openrouter, ollama, others
+    priority_order = ["nous", "openrouter", "ollama", "anthropic", "kimi-coding"]
+    
+    healthy = []
+    for provider in priority_order:
+        if provider in health_map and health_map[provider]["healthy"]:
+            healthy.append(provider)
+    
+    # Add any other healthy providers not in priority list
+    for provider, info in health_map.items():
+        if info["healthy"] and provider not in healthy:
+            healthy.append(provider)
+    
+    return healthy
+
+def update_profile_config(profile_name, new_provider):
+    """Update a profile's config to use a new provider."""
+    config_file = PROFILES_DIR / profile_name / "config.yaml"
+    
+    if not config_file.exists():
+        return False, "Config file not found"
+    
+    try:
+        with open(config_file) as f:
+            config = yaml.safe_load(f)
+        
+        # Update model provider
+        if "model" not in config:
+            config["model"] = {}
+        
+        old_provider = config["model"].get("provider", "unknown")
+        config["model"]["provider"] = new_provider
+        
+        # Update auxiliary providers if they were using the old provider
+        auxiliary = config.get("auxiliary", {})
+        for aux_name, aux_config in auxiliary.items():
+            if isinstance(aux_config, dict) and aux_config.get("provider") == old_provider:
+                aux_config["provider"] = new_provider
+        
+        # Write back
+        with open(config_file, "w") as f:
+            yaml.dump(config, f, default_flow_style=False)
+        
+        log("Updated " + profile_name + ": " + old_provider + " -> " + new_provider)
+        return True, "Updated"
+    
+    except Exception as e:
+        return False, str(e)
+
+def check_profiles(health_map):
+    """Check all profiles and update unhealthy providers."""
+    if not PROFILES_DIR.exists():
+        return
+    
+    fallback_providers = get_fallback_providers(health_map)
+    if not fallback_providers:
+        log("CRITICAL: No healthy providers available!")
+        return
+    
+    updated_profiles = []
+    
+    for profile_dir in PROFILES_DIR.iterdir():
+        if not profile_dir.is_dir():
+            continue
+        
+        profile_name = profile_dir.name
+        config_file = profile_dir / "config.yaml"
+        
+        if not config_file.exists():
+            continue
+        
+        try:
+            with open(config_file) as f:
+                config = yaml.safe_load(f)
+            
+            model_config = config.get("model", {})
+            if not isinstance(model_config, dict):
+                continue
+            
+            current_provider = model_config.get("provider", "")
+            if not current_provider:
+                continue
+            
+            # Check if current provider is healthy
+            if current_provider in health_map and health_map[current_provider]["healthy"]:
+                continue  # Provider is healthy, no action needed
+            
+            # Find best fallback
+            best_fallback = None
+            for provider in fallback_providers:
+                if provider != current_provider:
+                    best_fallback = provider
+                    break
+            
+            if not best_fallback:
+                log("No fallback for " + profile_name + " (current: " + current_provider + ")")
+                continue
+            
+            # Update profile
+            success, message = update_profile_config(profile_name, best_fallback)
+            if success:
+                updated_profiles.append({
+                    "profile": profile_name,
+                    "old_provider": current_provider,
+                    "new_provider": best_fallback
+                })
+        
+        except Exception as e:
+            log("Error processing " + profile_name + ": " + str(e))
+    
+    return updated_profiles
+
+def load_state():
+    """Load state from tmux-state.json."""
+    if STATE_FILE.exists():
+        try:
+            with open(STATE_FILE) as f:
+                return json.load(f)
+        except:
+            pass
+    return {}
+
+def save_state(state):
+    """Save state to tmux-state.json."""
+    LOG_DIR.mkdir(parents=True, exist_ok=True)
+    
+    with open(STATE_FILE, "w") as f:
+        json.dump(state, f, indent=2)
+
+def run_once():
+    """Run provider health check once."""
+    log("=== Provider Health Check ===")
+    
+    state = load_state()
+    
+    # Build health map
+    health_map = build_health_map()
+    
+    # Check profiles and update if needed
+    updated_profiles = check_profiles(health_map)
+    
+    # Update state
+    state["provider_health"] = health_map
+    state["last_provider_check"] = datetime.now(timezone.utc).isoformat()
+    
+    if updated_profiles:
+        state["last_profile_updates"] = updated_profiles
+    
+    save_state(state)
+    
+    # Summary
+    healthy_count = sum(1 for p in health_map.values() if p["healthy"])
+    total_count = len(health_map)
+    
+    log("Health: " + str(healthy_count) + "/" + str(total_count) + " providers healthy")
+    
+    if updated_profiles:
+        log("Updated " + str(len(updated_profiles)) + " profiles:")
+        for update in updated_profiles:
+            log("  " + update["profile"] + ": " + update["old_provider"] + " -> " + update["new_provider"])
+
+def show_status():
+    """Show provider health status."""
+    state = load_state()
+    health_map = state.get("provider_health", {})
+    
+    if not health_map:
+        print("No provider health data available. Run without --status first.")
+        return
+    
+    print("Provider Health (last updated: " + str(state.get("last_provider_check", "unknown")) + ")")
+    print("=" * 80)
+    
+    for provider, info in sorted(health_map.items()):
+        status = "HEALTHY" if info["healthy"] else "UNHEALTHY"
+        message = info.get("message", "")
+        api_key = "yes" if info.get("api_key_present") else "no"
+        
+        print(provider.ljust(20) + " " + status.ljust(10) + " API key: " + api_key + " - " + message)
+    
+    # Show recent updates
+    updates = state.get("last_profile_updates", [])
+    if updates:
+        print()
+        print("Recent Profile Updates:")
+        for update in updates:
+            print("  " + update["profile"] + ": " + update["old_provider"] + " -> " + update["new_provider"])
+
+def daemon_mode():
+    """Run continuously."""
+    log("Starting provider health daemon (check every 300s)")
+    
+    while True:
+        try:
+            run_once()
+            time.sleep(300)  # Check every 5 minutes
+        except KeyboardInterrupt:
+            log("Daemon stopped by user")
+            break
+        except Exception as e:
+            log("Error: " + str(e))
+            time.sleep(60)
+
+def main():
+    if "--status" in sys.argv:
+        show_status()
+    elif "--daemon" in sys.argv:
+        daemon_mode()
+    else:
+        run_once()
+
+if __name__ == "__main__":
+    main()

bin/quality-gate.py

@@ -0,0 +1,292 @@
+#!/usr/bin/env python3
+"""
+Quality Gate — Validate pipeline outputs before saving.
+
+Checks:
+- JSON schema validation for all output formats
+- Content quality (not empty, not duplicated, not toxic)
+- SOUL.md compliance for agent-facing content
+- Auto-reject bad outputs, re-queue for regeneration
+- Quality score tracking per pipeline
+
+Usage:
+    python3 quality-gate.py validate training-data/pairs.jsonl
+    python3 quality-gate.py validate --format training-pairs data.jsonl
+    python3 quality-gate.py score training-data/pairs.jsonl
+    python3 quality-gate.py stats
+"""
+
+import hashlib
+import json
+import os
+import sys
+from datetime import datetime, timezone
+from pathlib import Path
+
+HERMES_HOME = Path(os.environ.get("HERMES_HOME", os.path.expanduser("~/.hermes")))
+SCORE_FILE = HERMES_HOME / "quality-scores.jsonl"
+HASH_FILE = HERMES_HOME / "quality-hashes.json"
+MAX_HASH_AGE_DAYS = 7
+
+# ── Validators ─────────────────────────────────────────
+
+TOXIC_PATTERNS = [
+    "kill yourself", "kys", "you should die", "end it all",
+    "nobody loves you", "waste of life",
+]
+
+def validate_training_pair(entry):
+    """Validate a training pair (prompt + response)."""
+    errors = []
+    if not isinstance(entry, dict):
+        return ["Entry is not a dict"]
+    
+    prompt = entry.get("prompt", "") or entry.get("instruction", "") or ""
+    response = entry.get("response", "") or entry.get("output", "") or entry.get("completion", "") or ""
+    
+    if not prompt.strip():
+        errors.append("Empty prompt")
+    if not response.strip():
+        errors.append("Empty response")
+    if len(response) < 10:
+        errors.append(f"Response too short ({len(response)} chars)")
+    if len(prompt) > 10000:
+        errors.append(f"Prompt too long ({len(prompt)} chars)")
+    
+    # Toxicity check
+    combined = (prompt + " " + response).lower()
+    for pattern in TOXIC_PATTERNS:
+        if pattern in combined:
+            errors.append(f"Toxic content detected: '{pattern}'")
+    
+    return errors
+
+
+def validate_jsonl(filepath):
+    """Validate a JSONL file — each line must be valid JSON."""
+    errors = []
+    seen_hashes = set()
+    line_count = 0
+    
+    try:
+        with open(filepath) as f:
+            for i, line in enumerate(f, 1):
+                line = line.strip()
+                if not line:
+                    continue
+                line_count += 1
+                try:
+                    entry = json.loads(line)
+                except json.JSONDecodeError as e:
+                    errors.append(f"Line {i}: invalid JSON: {e}")
+                    continue
+                
+                # Duplicate detection
+                h = hashlib.sha256(line.encode()).hexdigest()[:16]
+                if h in seen_hashes:
+                    errors.append(f"Line {i}: duplicate content (hash {h})")
+                seen_hashes.add(h)
+                
+                # Content validation
+                if isinstance(entry, dict):
+                    pair_errors = validate_training_pair(entry)
+                    for pe in pair_errors:
+                        errors.append(f"Line {i}: {pe}")
+    
+    except Exception as e:
+        errors.append(f"File error: {e}")
+    
+    return errors, line_count, seen_hashes
+
+
+def validate_json(filepath):
+    """Validate a single JSON file."""
+    errors = []
+    try:
+        with open(filepath) as f:
+            data = json.load(f)
+    except json.JSONDecodeError as e:
+        return [f"Invalid JSON: {e}"], 0
+    
+    if isinstance(data, list):
+        seen = set()
+        for i, entry in enumerate(data):
+            if isinstance(entry, dict):
+                h = hashlib.sha256(json.dumps(entry, sort_keys=True).encode()).hexdigest()[:16]
+                if h in seen:
+                    errors.append(f"Index {i}: duplicate entry")
+                seen.add(h)
+    
+    return errors, len(data) if isinstance(data, list) else 1
+
+
+# ── Quality Scoring ────────────────────────────────────
+
+def score_file(filepath):
+    """Score a pipeline output file. Returns 0-100."""
+    path = Path(filepath)
+    if not path.exists():
+        return 0
+    
+    suffix = path.suffix.lower()
+    if suffix == ".jsonl":
+        errors, count, _ = validate_jsonl(filepath)
+    elif suffix == ".json":
+        errors, count = validate_json(filepath)
+    else:
+        return 50  # unknown format
+    
+    if count == 0:
+        return 0
+    
+    error_rate = len(errors) / count
+    score = max(0, int(100 * (1 - error_rate)))
+    
+    # Bonus for having content
+    if count >= 100:
+        score = min(100, score + 5)
+    
+    return score
+
+
+def record_score(filepath, score):
+    """Record quality score for tracking."""
+    HERMES_HOME.mkdir(parents=True, exist_ok=True)
+    entry = {
+        "timestamp": datetime.now(timezone.utc).isoformat(),
+        "file": str(filepath),
+        "score": score,
+    }
+    with open(SCORE_FILE, "a") as f:
+        f.write(json.dumps(entry) + "
+")
+
+
+# ── Dedup Hash Management ─────────────────────────────
+
+def load_hashes():
+    try:
+        return json.loads(HASH_FILE.read_text())
+    except Exception:
+        return {"entries": {}, "last_cleanup": None}
+
+
+def save_hashes(data):
+    HASH_FILE.parent.mkdir(parents=True, exist_ok=True)
+    HASH_FILE.write_text(json.dumps(data, indent=2))
+
+
+def cleanup_old_hashes(data, max_age_days=MAX_HASH_AGE_DAYS):
+    """Remove hash entries older than max_age_days."""
+    cutoff = datetime.now(timezone.utc).timestamp() - (max_age_days * 86400)
+    before = len(data.get("entries", {}))
+    data["entries"] = {
+        k: v for k, v in data.get("entries", {}).items()
+        if v.get("ts", 0) > cutoff
+    }
+    data["last_cleanup"] = datetime.now(timezone.utc).isoformat()
+    after = len(data["entries"])
+    return before - after
+
+
+# ── CLI ────────────────────────────────────────────────
+
+def cmd_validate(args):
+    filepath = args[0] if args else None
+    if not filepath or not os.path.exists(filepath):
+        print(f"ERROR: {filepath} not found")
+        sys.exit(1)
+    
+    suffix = Path(filepath).suffix.lower()
+    if suffix == ".jsonl":
+        errors, count, _ = validate_jsonl(filepath)
+    elif suffix == ".json":
+        errors, count = validate_json(filepath)
+    else:
+        print(f"Unsupported format: {suffix}")
+        sys.exit(1)
+    
+    score = score_file(filepath)
+    record_score(filepath, score)
+    
+    if errors:
+        for e in errors[:20]:
+            print(f"FAIL: {e}")
+        if len(errors) > 20:
+            print(f"... and {len(errors)-20} more")
+        print(f"
+Score: {score}/100 ({len(errors)} errors in {count} entries)")
+        sys.exit(1)
+    else:
+        print(f"OK: {filepath} ({count} entries, score {score}/100)")
+
+
+def cmd_score(args):
+    filepath = args[0] if args else None
+    if not filepath:
+        print("Usage: quality-gate.py score <file>")
+        sys.exit(1)
+    score = score_file(filepath)
+    print(f"Score: {score}/100")
+    record_score(filepath, score)
+
+
+def cmd_stats():
+    if not SCORE_FILE.exists():
+        print("No quality scores recorded yet.")
+        return
+    
+    scores = []
+    with open(SCORE_FILE) as f:
+        for line in f:
+            try:
+                scores.append(json.loads(line))
+            except Exception:
+                continue
+    
+    if not scores:
+        print("No scores recorded.")
+        return
+    
+    by_file = {}
+    for s in scores:
+        fname = s.get("file", "?")
+        by_file.setdefault(fname, []).append(s.get("score", 0))
+    
+    print("Quality Scores:")
+    for fname, scs in sorted(by_file.items()):
+        avg = sum(scs) / len(scs)
+        latest = scs[-1]
+        print(f"  {fname}: avg={avg:.0f}, latest={latest}, runs={len(scs)}")
+
+
+def cmd_cleanup():
+    data = load_hashes()
+    removed = cleanup_old_hashes(data)
+    save_hashes(data)
+    print(f"Cleaned up {removed} old hash entries (>{MAX_HASH_AGE_DAYS} days)")
+
+
+def main():
+    if len(sys.argv) < 2:
+        print("Usage: quality-gate.py <validate|score|stats|cleanup> [args]")
+        sys.exit(1)
+    
+    cmd = sys.argv[1]
+    args = sys.argv[2:]
+    
+    if cmd == "validate":
+        cmd_validate(args)
+    elif cmd == "score":
+        cmd_score(args)
+    elif cmd == "stats":
+        cmd_stats()
+    elif cmd == "cleanup":
+        cmd_cleanup()
+    else:
+        print(f"Unknown command: {cmd}")
+        sys.exit(1)
+
+
+if __name__ == "__main__":
+    main()

bin/soul_eval_gate.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 """
 Soul Eval Gate — The Conscience of the Training Pipeline
 

cron/jobs.json

@@ -196,7 +196,37 @@
       "paused_reason": null,
       "skills": [],
       "skill": null
+    },
+    {
+      "id": "tmux-supervisor-513",
+      "name": "Autonomous Cron Supervisor",
+      "prompt": "Load the tmux-supervisor skill and execute the monitoring protocol.\n\nCheck both `dev` and `timmy` tmux sessions for idle panes. Only send Telegram notifications on actionable events (idle, overflow, failure). Be silent when all agents are working.\n\nSteps:\n1. List all tmux sessions (skip 'Alexander')\n2. For each session, list windows and panes\n3. Capture each pane and classify state (idle vs active)\n4. For idle panes: read context, craft context-aware prompt\n5. Send /queue prompts to idle panes\n6. Verify prompts landed\n7. Only notify via Telegram if:\n   - A pane was prompted (idle detected)\n   - A pane shows context overflow (>80%)\n   - A pane is stuck or crashed\n8. If all panes are active: respond with [SILENT]",
+      "schedule": {
+        "kind": "interval",
+        "minutes": 7,
+        "display": "every 7m"
+      },
+      "schedule_display": "every 7m",
+      "repeat": {
+        "times": null,
+        "completed": 0
+      },
+      "enabled": true,
+      "created_at": "2026-04-15T03:00:00.000000+00:00",
+      "next_run_at": null,
+      "last_run_at": null,
+      "last_status": null,
+      "last_error": null,
+      "deliver": "telegram",
+      "origin": null,
+      "state": "scheduled",
+      "paused_at": null,
+      "paused_reason": null,
+      "skills": [
+        "tmux-supervisor"
+      ],
+      "skill": "tmux-supervisor"
     }
   ],
   "updated_at": "2026-04-13T02:00:00+00:00"
-}
+}

cron/pipeline-scheduler.yml

@@ -0,0 +1,9 @@
+- name: Nightly Pipeline Scheduler
+  schedule: '*/30 18-23,0-8 * * *' # Every 30 min, off-peak hours only
+  tasks:
+    - name: Check and start pipelines
+      shell: "bash scripts/nightly-pipeline-scheduler.sh"
+      env:
+        PIPELINE_TOKEN_LIMIT: "500000"
+        PIPELINE_PEAK_START: "9"
+        PIPELINE_PEAK_END: "18"

pipeline/quality_gate.py

@@ -0,0 +1,419 @@
+#!/usr/bin/env python3
+"""
+quality_gate.py — Quality Gate for Pipeline Outputs
+
+Validates all pipeline outputs before saving. Rejects bad outputs,
+tracks quality scores, and supports re-queue for regeneration.
+
+Usage:
+    python3 quality_gate.py --input output.jsonl --type training_pairs
+    python3 quality_gate.py --input output.jsonl --type knowledge
+    python3 quality_gate.py --input output.jsonl --type scene_descriptions
+    python3 quality_gate.py --dir pipeline/output/ --type training_pairs
+    python3 quality_gate.py --status  # show quality stats
+
+Exit codes:
+    0 = all outputs passed
+    1 = some outputs rejected
+    2 = file/parse error
+"""
+
+import json
+import os
+import sys
+import hashlib
+import re
+from pathlib import Path
+from datetime import datetime, timezone
+from dataclasses import dataclass, field, asdict
+from typing import List, Optional, Dict, Any
+
+STATS_FILE = Path.home() / ".hermes" / "pipeline" / "quality_stats.json"
+
+# --- Quality Check Types ---
+
+@dataclass
+class QualityResult:
+    """Result of a quality check on a single entry."""
+    passed: bool
+    checks_run: int
+    checks_failed: int
+    score: float  # 0.0-1.0
+    reasons: List[str] = field(default_factory=list)
+    entry_index: int = -1
+    hash: str = ""
+
+    def to_dict(self):
+        return asdict(self)
+
+
+@dataclass
+class GateReport:
+    """Report from a quality gate run."""
+    file: str
+    type: str
+    total: int
+    passed: int
+    rejected: int
+    score: float
+    rejected_indices: List[int] = field(default_factory=list)
+    timestamp: str = field(default_factory=lambda: datetime.now(timezone.utc).isoformat())
+
+    def to_dict(self):
+        return asdict(self)
+
+
+# ============================================================
+# Check Functions
+# ============================================================
+
+def entry_hash(entry: dict) -> str:
+    """Hash an entry for deduplication."""
+    return hashlib.sha256(json.dumps(entry, sort_keys=True, ensure_ascii=False).encode()).hexdigest()[:16]
+
+
+def check_not_empty(entry: dict, fields: List[str]) -> List[str]:
+    """Check that required fields are non-empty."""
+    errors = []
+    for f in fields:
+        val = entry.get(f)
+        if val is None:
+            errors.append(f"missing_field: {f}")
+        elif isinstance(val, str) and len(val.strip()) == 0:
+            errors.append(f"empty_field: {f}")
+        elif isinstance(val, list) and len(val) == 0:
+            errors.append(f"empty_list: {f}")
+    return errors
+
+
+def check_string_min_length(entry: dict, field_lengths: Dict[str, int]) -> List[str]:
+    """Check that string fields meet minimum lengths."""
+    errors = []
+    for f, min_len in field_lengths.items():
+        val = entry.get(f)
+        if isinstance(val, str) and len(val) < min_len:
+            errors.append(f"short_field: {f} ({len(val)} < {min_len})")
+    return errors
+
+
+def check_no_duplicates(entries: List[dict], key_fields: List[str]) -> Dict[int, List[str]]:
+    """Check for duplicate entries based on key fields."""
+    seen = {}
+    errors = {}
+    for i, entry in enumerate(entries):
+        key = tuple(entry.get(f, "") for f in key_fields)
+        key_str = str(key)
+        if key_str in seen:
+            errors[i] = [f"duplicate_of_index: {seen[key_str]}"]
+        else:
+            seen[key_str] = i
+    return errors
+
+
+def check_training_pair(entry: dict) -> List[str]:
+    """Validate a training pair (prompt/response)."""
+    errors = []
+    errors.extend(check_not_empty(entry, ["prompt", "response"]))
+
+    # Check response isn't just echoing the prompt
+    prompt = entry.get("prompt", "")
+    response = entry.get("response", "")
+    if prompt and response and prompt.strip() == response.strip():
+        errors.append("response_equals_prompt")
+
+    # Check response has substance
+    if isinstance(response, str) and len(response) < 10:
+        errors.append(f"response_too_short: {len(response)} chars")
+
+    return errors
+
+
+def check_scene_description(entry: dict) -> List[str]:
+    """Validate a scene description entry."""
+    errors = []
+    errors.extend(check_not_empty(entry, ["song", "beat", "lyric_line", "scene"]))
+
+    scene = entry.get("scene")
+    if isinstance(scene, dict):
+        errors.extend(check_not_empty(scene, ["mood", "colors", "composition", "camera", "description"]))
+        errors.extend(check_string_min_length(scene, {"description": 10}))
+
+        colors = scene.get("colors", [])
+        if isinstance(colors, list) and len(colors) > 5:
+            errors.append(f"too_many_colors: {len(colors)} > 5")
+
+    return errors
+
+
+def check_knowledge_entry(entry: dict) -> List[str]:
+    """Validate a knowledge file entry."""
+    errors = []
+    errors.extend(check_not_empty(entry, ["title", "content"]))
+
+    # Check for placeholder content
+    content = entry.get("content", "")
+    if isinstance(content, str):
+        placeholders = ["TODO", "FIXME", "PLACEHOLDER", "[INSERT", "lorem ipsum"]
+        for p in placeholders:
+            if p.lower() in content.lower():
+                errors.append(f"placeholder_content: '{p}' found")
+
+    errors.extend(check_string_min_length(entry, {"content": 50}))
+
+    return errors
+
+
+def check_prompt_enhancement(entry: dict) -> List[str]:
+    """Validate a prompt enhancement pair (terse/rich)."""
+    errors = []
+    errors.extend(check_not_empty(entry, ["terse", "rich"]))
+
+    terse = entry.get("terse", "")
+    rich = entry.get("rich", "")
+
+    # Rich should be longer than terse
+    if isinstance(terse, str) and isinstance(rich, str) and len(rich) <= len(terse):
+        errors.append("rich_not_longer_than_terse")
+
+    errors.extend(check_string_min_length(entry, {"rich": 20}))
+
+    return errors
+
+
+def check_adversary_entry(entry: dict) -> List[str]:
+    """Validate an adversary corpus entry."""
+    errors = []
+    errors.extend(check_not_empty(entry, ["id", "family", "prompt"]))
+
+    # Check prompt isn't empty or placeholder
+    prompt = entry.get("prompt", "")
+    if isinstance(prompt, str) and len(prompt) < 10:
+        errors.append(f"prompt_too_short: {len(prompt)} chars")
+
+    return errors
+
+
+def check_soul_compliance(text: str) -> List[str]:
+    """Check text for SOUL.md compliance (basic checks)."""
+    errors = []
+    text_lower = text.lower() if text else ""
+
+    # Check for content that violates core values
+    violations = [
+        ("compute.*value.*human.*life", "soul_violation: computing human life value"),
+        ("sexualiz.*child", "soul_violation: sexualizing children"),
+        ("create.*weapon.*kill", "soul_violation: weapon creation"),
+        ("enslav", "soul_violation: enslavement content"),
+    ]
+    for pattern, msg in violations:
+        if re.search(pattern, text_lower):
+            errors.append(msg)
+
+    return errors
+
+
+# ============================================================
+# Gate Runners
+# ============================================================
+
+CHECK_MAP = {
+    "training_pairs": check_training_pair,
+    "training_pair": check_training_pair,
+    "scene_descriptions": check_scene_description,
+    "scene_description": check_scene_description,
+    "knowledge": check_knowledge_entry,
+    "prompt_enhancement": check_prompt_enhancement,
+    "adversary": check_adversary_entry,
+    "adversary_corpus": check_adversary_entry,
+}
+
+
+def run_gate(input_path: str, entry_type: str) -> GateReport:
+    """Run quality gate on a JSONL file."""
+    path = Path(input_path)
+    if not path.exists():
+        return GateReport(file=str(path), type=entry_type, total=0, passed=0, rejected=0, score=0.0)
+
+    check_fn = CHECK_MAP.get(entry_type)
+    if not check_fn:
+        return GateReport(file=str(path), type=entry_type, total=0, passed=0, rejected=0, score=0.0,
+                          rejected_indices=[-1])  # unknown type
+
+    entries = []
+    with open(path) as f:
+        for line in f:
+            line = line.strip()
+            if line:
+                entries.append(json.loads(line))
+
+    # Deduplication check
+    key_fields = _get_key_fields(entry_type)
+    dup_errors = check_no_duplicates(entries, key_fields)
+
+    passed = 0
+    rejected = 0
+    rejected_indices = []
+    total_score = 0.0
+
+    for i, entry in enumerate(entries):
+        errors = check_fn(entry)
+
+        # Add duplicate errors
+        if i in dup_errors:
+            errors.extend(dup_errors[i])
+
+        # Add SOUL compliance check for text content
+        text_content = ""
+        for f in ["response", "rich", "description", "content", "lyric_line"]:
+            val = entry.get(f)
+            if isinstance(val, str):
+                text_content += val + " "
+        if isinstance(entry.get("scene"), dict):
+            text_content += entry["scene"].get("description", "")
+
+        soul_errors = check_soul_compliance(text_content)
+        errors.extend(soul_errors)
+
+        if errors:
+            rejected += 1
+            rejected_indices.append(i)
+        else:
+            passed += 1
+
+        # Score: 1.0 if no errors, decreasing with each error
+        entry_score = max(0.0, 1.0 - (len(errors) * 0.2))
+        total_score += entry_score
+
+    avg_score = total_score / len(entries) if entries else 0.0
+
+    report = GateReport(
+        file=str(path),
+        type=entry_type,
+        total=len(entries),
+        passed=passed,
+        rejected=rejected,
+        score=round(avg_score, 3),
+        rejected_indices=rejected_indices[:50],  # limit for readability
+    )
+
+    # Save stats
+    _save_stats(report)
+
+    return report
+
+
+def _get_key_fields(entry_type: str) -> List[str]:
+    """Get key fields for deduplication based on entry type."""
+    key_map = {
+        "training_pairs": ["prompt", "response"],
+        "training_pair": ["prompt", "response"],
+        "scene_descriptions": ["song", "beat"],
+        "scene_description": ["song", "beat"],
+        "knowledge": ["title"],
+        "prompt_enhancement": ["terse", "rich"],
+        "adversary": ["id", "prompt"],
+        "adversary_corpus": ["id", "prompt"],
+    }
+    return key_map.get(entry_type, ["id"])
+
+
+def _save_stats(report: GateReport):
+    """Append quality stats to the stats file."""
+    STATS_FILE.parent.mkdir(parents=True, exist_ok=True)
+
+    stats = []
+    if STATS_FILE.exists():
+        try:
+            with open(STATS_FILE) as f:
+                stats = json.load(f)
+        except (json.JSONDecodeError, IOError):
+            stats = []
+
+    stats.append(report.to_dict())
+
+    # Keep last 1000 entries
+    stats = stats[-1000:]
+
+    with open(STATS_FILE, "w") as f:
+        json.dump(stats, f, indent=2)
+
+
+def show_status():
+    """Show quality gate statistics."""
+    if not STATS_FILE.exists():
+        print("No quality stats found.")
+        return
+
+    with open(STATS_FILE) as f:
+        stats = json.load(f)
+
+    print(f"\nQuality Gate Stats — {len(stats)} runs")
+    print()
+
+    # Group by type
+    by_type = {}
+    for s in stats:
+        t = s.get("type", "unknown")
+        if t not in by_type:
+            by_type[t] = []
+        by_type[t].append(s)
+
+    for t, runs in sorted(by_type.items()):
+        total_entries = sum(r.get("total", 0) for r in runs)
+        total_passed = sum(r.get("passed", 0) for r in runs)
+        total_rejected = sum(r.get("rejected", 0) for r in runs)
+        avg_score = sum(r.get("score", 0) for r in runs) / len(runs) if runs else 0
+        print(f"  {t:25} {len(runs):4} runs | {total_entries:6} entries | {total_rejected:4} rejected | avg score: {avg_score:.3f}")
+
+
+def main():
+    import argparse
+    parser = argparse.ArgumentParser(description="Quality Gate for Pipeline Outputs")
+    parser.add_argument("--input", default=None, help="Input JSONL file")
+    parser.add_argument("--type", default=None, help="Entry type (training_pairs, scene_descriptions, knowledge, etc.)")
+    parser.add_argument("--dir", default=None, help="Process all JSONL files in directory")
+    parser.add_argument("--status", action="store_true", help="Show quality stats")
+    args = parser.parse_args()
+
+    if args.status:
+        show_status()
+        return
+
+    if args.dir:
+        for f in sorted(Path(args.dir).glob("*.jsonl")):
+            t = args.type or _infer_type(f.name)
+            report = run_gate(str(f), t)
+            _print_report(report)
+    elif args.input:
+        t = args.type or _infer_type(args.input)
+        report = run_gate(args.input, t)
+        _print_report(report)
+        sys.exit(0 if report.rejected == 0 else 1)
+    else:
+        parser.print_help()
+
+
+def _infer_type(filename: str) -> str:
+    """Infer entry type from filename."""
+    name = filename.lower()
+    if "scene" in name:
+        return "scene_descriptions"
+    if "training" in name or "pair" in name:
+        return "training_pairs"
+    if "knowledge" in name:
+        return "knowledge"
+    if "adversary" in name or "attack" in name:
+        return "adversary"
+    if "prompt" in name or "enhance" in name:
+        return "prompt_enhancement"
+    return "training_pairs"  # default
+
+
+def _print_report(report: GateReport):
+    """Print a human-readable gate report."""
+    status = "PASS" if report.rejected == 0 else f"FAIL ({report.rejected} rejected)"
+    print(f"  {report.file}: {status} | {report.passed}/{report.total} passed | score: {report.score:.3f}")
+
+
+if __name__ == "__main__":
+    main()

scripts/captcha_bypass_handler.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import json
 from hermes_tools import browser_navigate, browser_vision
 

scripts/diagram_meaning_extractor.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import json
 from hermes_tools import browser_navigate, browser_vision
 

scripts/nightly-pipeline-scheduler.md

@@ -0,0 +1,50 @@
+# Nightly Pipeline Scheduler
+
+Auto-starts batch pipelines when inference is available.
+
+## What It Does
+
+1. Checks inference provider health (OpenRouter, Ollama, RunPod)
+2. Checks if it's off-peak hours (configurable, default: after 6PM)
+3. Checks interactive session load (don't fight with live users)
+4. Checks daily token budget (configurable limit)
+5. Starts the highest-priority incomplete pipeline
+
+## Pipeline Priority Order
+
+| Priority | Pipeline | Deps | Max Tokens |
+|----------|----------|------|------------|
+| 1 | playground-factory | none | 100,000 |
+| 2 | training-factory | none | 150,000 |
+| 3 | knowledge-mine | training-factory running | 80,000 |
+| 4 | adversary | knowledge-mine running | 50,000 |
+| 5 | codebase-genome | none | 120,000 |
+
+## Usage
+
+```bash
+# Normal run (used by cron)
+./scripts/nightly-pipeline-scheduler.sh
+
+# Dry run (show what would start)
+./scripts/nightly-pipeline-scheduler.sh --dry-run
+
+# Status report
+./scripts/nightly-pipeline-scheduler.sh --status
+
+# Force start during peak hours
+./scripts/nightly-pipeline-scheduler.sh --force
+```
+
+## Configuration
+
+Set via environment variables:
+- `PIPELINE_TOKEN_LIMIT`: Daily token budget (default: 500,000)
+- `PIPELINE_PEAK_START`: Peak hours start (default: 9)
+- `PIPELINE_PEAK_END`: Peak hours end (default: 18)
+- `HERMES_HOME`: Hermes home directory (default: ~/.hermes)
+
+## Cron
+
+Runs every 30 minutes. Off-peak only (unless --force).
+See `cron/pipeline-scheduler.yml`.

scripts/nightly-pipeline-scheduler.sh

@@ -0,0 +1,383 @@
+#!/usr/bin/env bash
+# nightly-pipeline-scheduler.sh — Auto-start batch pipelines when inference is available.
+#
+# Checks provider health, pipeline progress, token budget, and interactive load.
+# Starts the highest-priority incomplete pipeline that can run.
+#
+# Usage:
+#   ./scripts/nightly-pipeline-scheduler.sh          # Normal run
+#   ./scripts/nightly-pipeline-scheduler.sh --dry-run # Show what would start
+#   ./scripts/nightly-pipeline-scheduler.sh --status  # Pipeline status report
+
+set -euo pipefail
+
+# --- Configuration ---
+HERMES_HOME="${HERMES_HOME:-$HOME/.hermes}"
+BUDGET_FILE="${HERMES_HOME}/pipeline_budget.json"
+STATE_FILE="${HERMES_HOME}/pipeline_state.json"
+LOG_FILE="${HERMES_HOME}/logs/pipeline-scheduler.log"
+TOKEN_DAILY_LIMIT="${PIPELINE_TOKEN_LIMIT:-500000}"
+PEAK_HOURS_START="${PIPELINE_PEAK_START:-9}"
+PEAK_HOURS_END="${PIPELINE_PEAK_END:-18}"
+
+# Pipeline definitions (priority order)
+# Each pipeline: name, script, max_tokens, dependencies
+PIPELINES=(
+    "playground-factory|scripts/pipeline_playground_factory.sh|100000|none"
+    "training-factory|scripts/pipeline_training_factory.sh|150000|none"
+    "knowledge-mine|scripts/pipeline_knowledge_mine.sh|80000|training-factory"
+    "adversary|scripts/pipeline_adversary.sh|50000|knowledge-mine"
+    "codebase-genome|scripts/pipeline_codebase_genome.sh|120000|none"
+)
+
+# --- Colors ---
+RED='\033[0;31m'
+GREEN='\033[0;32m'
+YELLOW='\033[0;33m'
+CYAN='\033[0;36m'
+NC='\033[0m'
+
+# --- Helpers ---
+now_hour() { date +%-H; }
+is_peak_hours() {
+    local h=$(now_hour)
+    [[ $h -ge $PEAK_HOURS_START && $h -lt $PEAK_HOURS_END ]]
+}
+
+ensure_dirs() {
+    mkdir -p "$(dirname "$LOG_FILE")" "$(dirname "$BUDGET_FILE")" "$(dirname "$STATE_FILE")"
+}
+
+log() { echo "[$(date '+%Y-%m-%d %H:%M:%S')] $*" | tee -a "$LOG_FILE"; }
+
+get_budget_used_today() {
+    if [[ -f "$BUDGET_FILE" ]]; then
+        local today=$(date +%Y-%m-%d)
+        python3 -c "
+import json, sys
+with open('$BUDGET_FILE') as f:
+    d = json.load(f)
+print(d.get('daily', {}).get('$today', {}).get('tokens_used', 0))
+" 2>/dev/null || echo 0
+    else
+        echo 0
+    fi
+}
+
+get_budget_remaining() {
+    local used=$(get_budget_used_today)
+    echo $((TOKEN_DAILY_LIMIT - used))
+}
+
+update_budget() {
+    local pipeline="$1"
+    local tokens="$2"
+    local today=$(date +%Y-%m-%d)
+    python3 -c "
+import json, os
+path = '$BUDGET_FILE'
+d = {}
+if os.path.exists(path):
+    with open(path) as f:
+        d = json.load(f)
+daily = d.setdefault('daily', {})
+day = daily.setdefault('$today', {'tokens_used': 0, 'pipelines': {}})
+day['tokens_used'] = day.get('tokens_used', 0) + $tokens
+day['pipelines']['$pipeline'] = day['pipelines'].get('$pipeline', 0) + $tokens
+with open(path, 'w') as f:
+    json.dump(d, f, indent=2)
+"
+}
+
+get_pipeline_state() {
+    if [[ -f "$STATE_FILE" ]]; then
+        cat "$STATE_FILE"
+    else
+        echo "{}"
+    fi
+}
+
+set_pipeline_state() {
+    local pipeline="$1"
+    local state="$2"  # running, complete, failed, skipped
+    python3 -c "
+import json, os
+path = '$STATE_FILE'
+d = {}
+if os.path.exists(path):
+    with open(path) as f:
+        d = json.load(f)
+d['$pipeline'] = {'state': '$state', 'updated': '$(date -Iseconds)'}
+with open(path, 'w') as f:
+    json.dump(d, f, indent=2)
+"
+}
+
+is_pipeline_complete() {
+    local pipeline="$1"
+    python3 -c "
+import json, os
+path = '$STATE_FILE'
+if not os.path.exists(path):
+    print('false')
+else:
+    with open(path) as f:
+        d = json.load(f)
+    state = d.get('$pipeline', {}).get('state', 'not_started')
+    print('true' if state == 'complete' else 'false')
+" 2>/dev/null || echo false
+}
+
+is_pipeline_running() {
+    local pipeline="$1"
+    python3 -c "
+import json, os
+path = '$STATE_FILE'
+if not os.path.exists(path):
+    print('false')
+else:
+    with open(path) as f:
+        d = json.load(f)
+    state = d.get('$pipeline', {}).get('state', 'not_started')
+    print('true' if state == 'running' else 'false')
+" 2>/dev/null || echo false
+}
+
+check_dependency() {
+    local dep="$1"
+    if [[ "$dep" == "none" ]]; then
+        return 0
+    fi
+    # For knowledge-mine: training-factory must be running or complete
+    if [[ "$dep" == "training-factory" ]]; then
+        local state=$(python3 -c "
+import json, os
+path = '$STATE_FILE'
+if not os.path.exists(path):
+    print('not_started')
+else:
+    with open(path) as f:
+        d = json.load(f)
+    print(d.get('training-factory', {}).get('state', 'not_started'))
+" 2>/dev/null || echo "not_started")
+        [[ "$state" == "running" || "$state" == "complete" ]]
+        return $?
+    fi
+    # For adversary: knowledge-mine must be at least 50% done
+    # Simplified: check if it's running (we'd need progress tracking for 50%)
+    if [[ "$dep" == "knowledge-mine" ]]; then
+        local state=$(python3 -c "
+import json, os
+path = '$STATE_FILE'
+if not os.path.exists(path):
+    print('not_started')
+else:
+    with open(path) as f:
+        d = json.load(f)
+    print(d.get('knowledge-mine', {}).get('state', 'not_started'))
+" 2>/dev/null || echo "not_started")
+        [[ "$state" == "running" || "$state" == "complete" ]]
+        return $?
+    fi
+    return 0
+}
+
+check_inference_available() {
+    # Check if any inference provider is responding
+    # 1. Check OpenRouter
+    local or_ok=$(curl -s -o /dev/null -w "%{http_code}" \
+        --connect-timeout 5 "https://openrouter.ai/api/v1/models" 2>/dev/null || echo "000")
+
+    # 2. Check local Ollama
+    local ollama_ok=$(curl -s -o /dev/null -w "%{http_code}" \
+        --connect-timeout 5 "http://localhost:11434/api/tags" 2>/dev/null || echo "000")
+
+    # 3. Check RunPod (if configured)
+    local runpod_ok="000"
+    if [[ -n "${RUNPOD_ENDPOINT:-}" ]]; then
+        runpod_ok=$(curl -s -o /dev/null -w "%{http_code}" \
+            --connect-timeout 5 "$RUNPOD_ENDPOINT/health" 2>/dev/null || echo "000")
+    fi
+
+    if [[ "$or_ok" == "200" || "$ollama_ok" == "200" || "$runpod_ok" == "200" ]]; then
+        return 0
+    fi
+    return 1
+}
+
+check_interactive_load() {
+    # Check if there are active interactive sessions (don't fight with live users)
+    # Look for tmux panes with active hermes sessions
+    local active=$(tmux list-panes -a -F '#{pane_pid} #{pane_current_command}' 2>/dev/null \
+        | grep -c "hermes\|python3" || echo 0)
+
+    # If more than 3 interactive sessions, skip pipeline start
+    if [[ $active -gt 3 ]]; then
+        return 1
+    fi
+    return 0
+}
+
+start_pipeline() {
+    local name="$1"
+    local script="$2"
+    local max_tokens="$3"
+    local budget_remaining="$4"
+    local mode="${5:-run}"
+
+    if [[ "$budget_remaining" -lt "$max_tokens" ]]; then
+        log "SKIP $name: insufficient budget ($budget_remaining < $max_tokens tokens)"
+        return 1
+    fi
+
+    if [[ ! -f "$script" ]]; then
+        log "SKIP $name: script not found ($script)"
+        return 1
+    fi
+
+    if [[ "$mode" == "dry-run" ]]; then
+        log "DRY-RUN: Would start $name (budget: $budget_remaining, needs: $max_tokens)"
+        return 0
+    fi
+
+    log "START $name (budget: $budget_remaining, max_tokens: $max_tokens)"
+    set_pipeline_state "$name" "running"
+
+    # Run in background, capture output
+    local log_path="${HERMES_HOME}/logs/pipeline-${name}.log"
+    bash "$script" --max-tokens "$max_tokens" >> "$log_path" 2>&1 &
+    local pid=$!
+
+    # Wait a moment to check if it started OK
+    sleep 2
+    if kill -0 $pid 2>/dev/null; then
+        log "RUNNING $name (PID: $pid, log: $log_path)"
+        # Record the PID
+        python3 -c "
+import json, os
+path = '$STATE_FILE'
+d = {}
+if os.path.exists(path):
+    with open(path) as f:
+        d = json.load(f)
+d['$name']['pid'] = $pid
+with open(path, 'w') as f:
+    json.dump(d, f, indent=2)
+"
+        return 0
+    else
+        log "FAIL $name: script exited immediately"
+        set_pipeline_state "$name" "failed"
+        return 1
+    fi
+}
+
+# --- Main ---
+main() {
+    local mode="${1:-run}"
+    ensure_dirs
+
+    log "=== Pipeline Scheduler ($mode) ==="
+
+    # Check 1: Is inference available?
+    if ! check_inference_available; then
+        log "No inference provider available. Skipping all pipelines."
+        exit 0
+    fi
+    log "Inference: AVAILABLE"
+
+    # Check 2: Is it peak hours?
+    if is_peak_hours && [[ "$mode" != "--force" ]]; then
+        local h=$(now_hour)
+        log "Peak hours ($h:00). Skipping pipeline start. Use --force to override."
+        exit 0
+    fi
+    log "Off-peak: OK"
+
+    # Check 3: Interactive load
+    if ! check_interactive_load && [[ "$mode" != "--force" ]]; then
+        log "High interactive load. Skipping pipeline start."
+        exit 0
+    fi
+    log "Interactive load: OK"
+
+    # Check 4: Token budget
+    local budget=$(get_budget_remaining)
+    log "Token budget remaining: $budget / $TOKEN_DAILY_LIMIT"
+
+    if [[ $budget -le 0 ]]; then
+        log "Daily token budget exhausted. Stopping."
+        exit 0
+    fi
+
+    # Check 5: Pipeline status
+    if [[ "$mode" == "--status" ]]; then
+        echo -e "${CYAN}Pipeline Status:${NC}"
+        echo "────────────────────────────────────────────────────"
+        for entry in "${PIPELINES[@]}"; do
+            IFS='|' read -r name script max_tokens dep <<< "$entry"
+            local state=$(python3 -c "
+import json, os
+path = '$STATE_FILE'
+if not os.path.exists(path):
+    print('not_started')
+else:
+    with open(path) as f:
+        d = json.load(f)
+    print(d.get('$name', {}).get('state', 'not_started'))
+" 2>/dev/null || echo "not_started")
+
+            local color=$NC
+            case "$state" in
+                running)  color=$YELLOW ;;
+                complete) color=$GREEN ;;
+                failed)   color=$RED ;;
+            esac
+            printf "  %-25s %b%s%b (max: %s tokens, dep: %s)\n" "$name" "$color" "$state" "$NC" "$max_tokens" "$dep"
+        done
+        echo "────────────────────────────────────────────────────"
+        echo "  Budget: $budget / $TOKEN_DAILY_LIMIT tokens remaining"
+        echo "  Peak hours: $PEAK_HOURS_START:00 - $PEAK_HOURS_END:00"
+        exit 0
+    fi
+
+    # Find and start the highest-priority incomplete pipeline
+    local started=0
+    for entry in "${PIPELINES[@]}"; do
+        IFS='|' read -r name script max_tokens dep <<< "$entry"
+
+        # Skip if already running or complete
+        if [[ "$(is_pipeline_running $name)" == "true" ]]; then
+            log "SKIP $name: already running"
+            continue
+        fi
+        if [[ "$(is_pipeline_complete $name)" == "true" ]]; then
+            log "SKIP $name: already complete"
+            continue
+        fi
+
+        # Check dependency
+        if ! check_dependency "$dep"; then
+            log "SKIP $name: dependency $dep not met"
+            continue
+        fi
+
+        # Try to start
+        if start_pipeline "$name" "$script" "$max_tokens" "$budget" "$mode"; then
+            started=1
+            # Only start one pipeline per run (let it claim tokens before next check)
+            # Exception: playground-factory and training-factory can run in parallel
+            if [[ "$name" != "playground-factory" && "$name" != "training-factory" ]]; then
+                break
+            fi
+        fi
+    done
+
+    if [[ $started -eq 0 ]]; then
+        log "No pipelines to start (all complete, running, or blocked)."
+    fi
+
+    log "=== Pipeline Scheduler done ==="
+}
+
+main "$@"

scripts/visual_pr_reviewer.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import json
 from hermes_tools import browser_navigate, browser_vision
 

tests/test_glitch_detector.py

@@ -19,9 +19,11 @@ from glitch_patterns import (
     GlitchPattern,
     GlitchSeverity,
     MATRIX_GLITCH_PATTERNS,
+    THREEJS_CATEGORIES,
     build_vision_prompt,
     get_pattern_by_category,
     get_patterns_by_severity,
+    get_threejs_patterns,
 )
 
 from matrix_glitch_detector import (
@@ -40,7 +42,7 @@ class TestGlitchPatterns(unittest.TestCase):
 
     def test_pattern_count(self):
         """Verify we have a reasonable number of defined patterns."""
-        self.assertGreaterEqual(len(MATRIX_GLITCH_PATTERNS), 8)
+        self.assertGreaterEqual(len(MATRIX_GLITCH_PATTERNS), 14)  # 10 generic + 6 Three.js
 
     def test_all_patterns_have_required_fields(self):
         """Every pattern must have category, name, description, severity, prompts."""
@@ -88,6 +90,9 @@ class TestGlitchPatterns(unittest.TestCase):
         self.assertIn("Floating Object", prompt)
         self.assertIn("Z-Fighting", prompt)
         self.assertIn("Missing", prompt)
+        # Three.js patterns should be included
+        self.assertIn("Shader Compilation Failure", prompt)
+        self.assertIn("Bloom Overflow", prompt)
 
     def test_build_vision_prompt_subset(self):
         """Vision prompt with subset should only include specified patterns."""
@@ -248,7 +253,7 @@ class TestGlitchDetector(unittest.TestCase):
 
         try:
             report = run_demo(output_path)
-            self.assertEqual(len(report.glitches), 4)
+            self.assertEqual(len(report.glitches), 6)  # 4 original + 2 Three.js
             self.assertGreater(report.summary["total_glitches"], 0)
             self.assertTrue(output_path.exists())
 
@@ -260,6 +265,93 @@ class TestGlitchDetector(unittest.TestCase):
             output_path.unlink(missing_ok=True)
 
 
+class TestThreeJsPatterns(unittest.TestCase):
+    """Tests for Three.js-specific glitch patterns (timmy-config#543)."""
+
+    def test_get_threejs_patterns_returns_only_threejs(self):
+        """get_threejs_patterns() should return only Three.js categories."""
+        patterns = get_threejs_patterns()
+        self.assertEqual(len(patterns), 6)
+        for p in patterns:
+            self.assertIn(p.category, THREEJS_CATEGORIES)
+
+    def test_threejs_patterns_have_required_fields(self):
+        """All Three.js patterns must have valid fields."""
+        for p in get_threejs_patterns():
+            self.assertIsInstance(p.category, GlitchCategory)
+            self.assertTrue(p.name)
+            self.assertTrue(p.description)
+            self.assertIsInstance(p.severity, GlitchSeverity)
+            self.assertGreater(len(p.detection_prompts), 0)
+            self.assertGreater(len(p.visual_indicators), 0)
+
+    def test_shader_failure_is_critical(self):
+        """Shader compilation failure should be CRITICAL severity."""
+        p = get_pattern_by_category(GlitchCategory.SHADER_FAILURE)
+        self.assertIsNotNone(p)
+        self.assertEqual(p.severity, GlitchSeverity.CRITICAL)
+
+    def test_texture_placeholder_is_critical(self):
+        """Texture placeholder (1x1 white) should be CRITICAL severity."""
+        p = get_pattern_by_category(GlitchCategory.TEXTURE_PLACEHOLDER)
+        self.assertIsNotNone(p)
+        self.assertEqual(p.severity, GlitchSeverity.CRITICAL)
+
+    def test_infer_severity_shader_failure(self):
+        """Shader failure should infer critical/high."""
+        self.assertEqual(_infer_severity("shader_failure", 0.8), "critical")
+        self.assertEqual(_infer_severity("shader_failure", 0.5), "high")
+
+    def test_infer_severity_texture_placeholder(self):
+        """Texture placeholder should infer critical/high."""
+        self.assertEqual(_infer_severity("texture_placeholder", 0.8), "critical")
+        self.assertEqual(_infer_severity("texture_placeholder", 0.5), "high")
+
+    def test_infer_severity_uv_mapping(self):
+        """UV mapping error should infer high/medium."""
+        self.assertEqual(_infer_severity("uv_mapping_error", 0.8), "high")
+        self.assertEqual(_infer_severity("uv_mapping_error", 0.5), "medium")
+
+    def test_infer_severity_frustum_culling(self):
+        """Frustum culling should infer medium/low."""
+        self.assertEqual(_infer_severity("frustum_culling", 0.7), "medium")
+        self.assertEqual(_infer_severity("frustum_culling", 0.4), "low")
+
+    def test_infer_severity_shadow_map(self):
+        """Shadow map artifact should infer medium/low."""
+        self.assertEqual(_infer_severity("shadow_map_artifact", 0.7), "medium")
+        self.assertEqual(_infer_severity("shadow_map_artifact", 0.4), "low")
+
+    def test_infer_severity_bloom_overflow(self):
+        """Bloom overflow should infer medium/low (default path)."""
+        self.assertEqual(_infer_severity("bloom_overflow", 0.7), "medium")
+        self.assertEqual(_infer_severity("bloom_overflow", 0.4), "low")
+
+    def test_threejs_patterns_in_vision_prompt(self):
+        """Three.js patterns should appear in the composite vision prompt."""
+        prompt = build_vision_prompt()
+        self.assertIn("shader_failure", prompt)
+        self.assertIn("texture_placeholder", prompt)
+        self.assertIn("uv_mapping_error", prompt)
+        self.assertIn("frustum_culling", prompt)
+        self.assertIn("shadow_map_artifact", prompt)
+        self.assertIn("bloom_overflow", prompt)
+
+    def test_threejs_subset_prompt(self):
+        """Building prompt from Three.js-only patterns should work."""
+        threejs = get_threejs_patterns()
+        prompt = build_vision_prompt(threejs)
+        self.assertIn("Shader Compilation Failure", prompt)
+        self.assertNotIn("Floating Object", prompt)  # generic, not Three.js
+
+    def test_report_metadata_version(self):
+        """Report metadata should reference both issues."""
+        report = run_demo()
+        self.assertEqual(report.metadata["detector_version"], "0.2.0")
+        self.assertIn("543", report.metadata["reference"])
+
+
+
 class TestIntegration(unittest.TestCase):
     """Integration-level tests."""
 
@@ -276,6 +368,13 @@ class TestIntegration(unittest.TestCase):
         expected = {"floating_assets", "z_fighting", "missing_textures", "clipping", "broken_normals"}
         self.assertTrue(expected.issubset(category_values))
 
+    def test_patterns_cover_threejs_themes(self):
+        """Patterns should cover Three.js-specific glitch themes (#543)."""
+        category_values = {p.category.value for p in MATRIX_GLITCH_PATTERNS}
+        threejs_expected = {"shader_failure", "texture_placeholder", "uv_mapping_error",
+                           "frustum_culling", "shadow_map_artifact", "bloom_overflow"}
+        self.assertTrue(threejs_expected.issubset(category_values))
+
 
 if __name__ == "__main__":
     unittest.main()

training/scripts/augment_pairs.py

@@ -0,0 +1,129 @@
+#!/usr/bin/env python3
+"""
+augment_pairs.py — Training data augmentation: paraphrase and translate.
+
+Usage:
+    python3 augment_pairs.py --input data.jsonl
+    python3 augment_pairs.py --input data.jsonl --paraphrases 3 --langs es,fr,de
+    python3 augment_pairs.py --input data.jsonl --llm-endpoint http://localhost:11434/v1
+"""
+
+import json, os, sys, re, random
+from pathlib import Path
+
+random.seed(42)
+
+PARAPHRASE_TRANSFORMS = [
+    lambda s: re.sub(r"(\w+), (\w+)", r"\2, \1", s, count=1),
+    lambda s: f"A beautifully rendered scene: {s[0].lower()}{s[1:]}" if len(s) > 10 else s,
+    lambda s: s.replace("A ", "The ").replace("An ", "The ") if s.startswith(("A ", "An ")) else f"Here, {s[0].lower()}{s[1:]}",
+    lambda s: f"In a cinematic frame: {s}" if len(s) > 20 else s,
+    lambda s: s if ", " not in s else ", ".join(s.split(", ")[:2]),
+]
+
+TRANSLATIONS = {
+    "es": {"the":"el","a":"un","is":"es","in":"en","of":"de","and":"y","with":"con","scene":"escena","light":"luz","dark":"oscuro","warm":"cálido","rain":"lluvia","sun":"sol","moon":"luna","sky":"cielo","forest":"bosque","mountain":"montaña","ocean":"océano","golden":"dorado","blue":"azul","red":"rojo","green":"verde","silence":"silencio","dream":"sueño","love":"amor","hope":"esperanza","fear":"miedo","joy":"alegría","peace":"paz","beautiful":"hermoso","sad":"triste","shadow":"sombra","color":"color","silver":"plateado","white":"blanco","black":"negro","portray":"retrato"},
+    "fr": {"the":"le","a":"un","is":"est","in":"dans","of":"de","and":"et","with":"avec","scene":"scène","light":"lumière","dark":"sombre","warm":"chaud","rain":"pluie","sun":"soleil","moon":"lune","sky":"ciel","forest":"forêt","mountain":"montagne","ocean":"océan","golden":"doré","blue":"bleu","red":"rouge","green":"vert","silence":"silence","dream":"rêve","love":"amour","hope":"espoir","fear":"peur","joy":"joie","peace":"paix","beautiful":"beau","sad":"triste","shadow":"ombre","color":"couleur","silver":"argenté","white":"blanc","black":"noir"},
+    "de": {"the":"der","a":"ein","is":"ist","in":"in","of":"von","and":"und","with":"mit","scene":"Szene","light":"Licht","dark":"dunkel","warm":"warm","rain":"Regen","sun":"Sonne","moon":"Mond","sky":"Himmel","forest":"Wald","mountain":"Berg","ocean":"Ozean","golden":"golden","blue":"blau","red":"rot","green":"grün","silence":"Stille","dream":"Traum","love":"Liebe","hope":"Hoffnung","fear":"Angst","joy":"Freude","peace":"Frieden","beautiful":"schön","sad":"traurig","shadow":"Schatten","color":"Farbe","silver":"silbern","white":"weiß","black":"schwarz"},
+}
+
+LANG_NAMES = {"es": "Spanish", "fr": "French", "de": "German"}
+
+
+def detect_text_field(entry):
+    for f in ["rich","terse","text","content","lyric_line","description","scene_description","prompt","scene"]:
+        if f in entry and isinstance(entry[f], str) and len(entry[f]) > 5:
+            return f
+    for k, v in entry.items():
+        if isinstance(v, str) and len(v) > 5:
+            return k
+    return None
+
+
+def paraphrase(text):
+    t = random.choice(PARAPHRASE_TRANSFORMS)(text)
+    if t == text:
+        t = text.replace(" and ", " & ").replace(" with ", " alongside ")
+        if t == text:
+            t = f"In this scene: {text[0].lower()}{text[1:]}" if text[0].isupper() else text
+    return t
+
+
+def translate(text, lang):
+    d = TRANSLATIONS.get(lang, {})
+    words = text.split()
+    out = []
+    for w in words:
+        lo = w.lower().strip(".,;:!?")
+        suf = w[len(w.rstrip(".,;:!?")):]
+        if lo in d:
+            out.append(d[lo] + suf)
+        else:
+            out.append(w)
+    return " ".join(out)
+
+
+def augment_file(input_path, output_path=None, n_para=3, langs=None, llm_endpoint=None):
+    input_path = Path(input_path)
+    if output_path is None:
+        output_path = input_path.parent / f"{input_path.stem}_augmented{input_path.suffix}"
+
+    entries = [json.loads(l) for l in open(input_path) if l.strip()]
+    if not entries:
+        print(f"No entries in {input_path}"); return 0
+
+    tf = detect_text_field(entries[0])
+    if not tf:
+        print(f"ERROR: No text field in {input_path}", file=sys.stderr); return 0
+
+    print(f"Input: {input_path} ({len(entries)} entries, field={tf})")
+
+    aug_count = 0
+    with open(output_path, "w") as out:
+        for e in entries:
+            out.write(json.dumps(e, ensure_ascii=False) + "\n")
+        for i, e in enumerate(entries):
+            text = e[tf]
+            # Paraphrases
+            for p in range(n_para):
+                para = paraphrase(text)
+                if para != text:
+                    ne = dict(e); ne[tf] = para
+                    ne["_augmentation"] = f"paraphrase_{p+1}"
+                    ne["_original"] = text[:100]
+                    out.write(json.dumps(ne, ensure_ascii=False) + "\n")
+                    aug_count += 1
+            # Translations
+            for lang in (langs or []):
+                tr = translate(text, lang)
+                if tr != text:
+                    ne = dict(e); ne[tf] = tr
+                    ne["_augmentation"] = f"translate_{lang}"
+                    ne["_language"] = lang
+                    ne["_original"] = text[:100]
+                    out.write(json.dumps(ne, ensure_ascii=False) + "\n")
+                    aug_count += 1
+            if (i+1) % 100 == 0:
+                print(f"  {i+1}/{len(entries)} done ({aug_count} augmented)")
+
+    total = len(entries) + aug_count
+    print(f"Done: {len(entries)} originals + {aug_count} augmented = {total}")
+    print(f"Output: {output_path}")
+    return aug_count
+
+
+def main():
+    import argparse
+    p = argparse.ArgumentParser()
+    p.add_argument("--input", required=True)
+    p.add_argument("--output", default=None)
+    p.add_argument("--paraphrases", type=int, default=3)
+    p.add_argument("--langs", default="es,fr,de")
+    p.add_argument("--llm-endpoint", default=None)
+    args = p.parse_args()
+    langs = [l.strip() for l in args.langs.split(",") if l.strip()] if args.langs else []
+    augment_file(args.input, args.output, args.paraphrases, langs, args.llm_endpoint)
+
+
+if __name__ == "__main__":
+    main()