#!/usr/bin/env python3
"""Dialogue audit for talking-head cuts — the four defects TJ had to catch by hand on 2026-09-05.

usage: dialogue_audit.py <cuts.json> <words.json> [--parts DIR] [--fix-out cuts_fixed.json]

cuts.json  : list of [section, id, in, out] in seconds on the SAME clock as words.json
             (in may be a string "a–b + c–d" for already-spliced parts; those are audited by their rendered part only)
words.json : list of [start, end, word] — word-level whisper output, one continuous clock
--parts    : directory of rendered parts <id>.mp4 — transcribes the first and last 2.5 s of each and
             checks the first/last expected word is present (catches clipped words AND bloopers at the head)

Checks
 1 OUT clips the last word   : out < last word end + 0.15
 2 IN inside a word          : a word starts before `in` and ends after it
 3 mid-sentence pause        : consecutive word STARTS > 1.5 s apart with no sentence punctuation before
                               (whisper word END times stretch across pauses — never use them for gaps)
 4 blooper at the head       : the first two words recur within 4 s of the in-point (a restart)
 5 head/tail transcription   : with --parts, whisper-small on the first/last 2.5 s of the rendered part
With --fix-out, writes corrected in/out points: out = last word end + 0.25 (capped 0.06 before the next word),
in = first word start − 0.10 (or the midpoint if the previous word abuts). Pauses and bloopers are reported,
not auto-fixed — cutting them is a jump cut and TJ decides.
"""
import json, sys, re, subprocess, os, tempfile
def norm(s): return re.sub(r"[^a-z0-9]","",s.lower())
def main():
    a=sys.argv[1:]; cuts=json.load(open(a[0])); words=[tuple(w) for w in json.load(open(a[1]))]; words.sort()
    parts=a[a.index("--parts")+1] if "--parts" in a else None; fixout=a[a.index("--fix-out")+1] if "--fix-out" in a else None
    fixed=[]; nflag=0
    for c in cuts:
        sec,id_,ci,co=c[:4]
        if isinstance(ci,str) or not isinstance(co,(int,float)): fixed.append(c); continue
        ci=float(ci); co=float(co); flags=[]
        ws=[w for w in words if ci-0.05<=w[0]<co]
        if not ws: fixed.append(c); continue
        first=ws[0]; last=[w for w in words if w[0]<co][-1]; nxt=[w for w in words if w[0]>=co]; nxt=nxt[0] if nxt else None
        if co<last[1]+0.15: flags.append(f"OUT clips '{last[2]}' (ends {last[1]:.2f}, cut {co:.2f})")
        inside=[w for w in words if w[0]<ci<w[1]-0.03]
        if inside: flags.append(f"IN inside '{inside[0][2]}' ({inside[0][0]:.2f}-{inside[0][1]:.2f})")
        for w1,w2 in zip(ws,ws[1:]):
            g=w2[0]-w1[0]
            if (g>1.5 and not w1[2].endswith((".","?","!"))) or g>2.4: flags.append(f"PAUSE {g:.1f}s '{w1[2]}'→'{w2[2]}' at {w1[0]:.2f}")
        head=[norm(w[2]) for w in ws if w[0]<ci+4.0]
        if len(head)>=4:
            bg=head[0]+head[1]
            for k in range(1,len(head)-1):
                if head[k]+head[k+1]==bg: flags.append(f"BLOOPER: '{ws[0][2]} {ws[1][2]}' repeats at {ws[k][0]:.2f} — start there")
        nb=min(last[1]+0.25, nxt[0]-0.06) if nxt else last[1]+0.25; nb=max(nb,co)
        prev=[w for w in words if w[1]<=first[0]]; prev=prev[-1] if prev else None
        na=first[0]-0.10
        if prev and na<prev[1]: na=(prev[1]+first[0])/2
        if parts:
            p=os.path.join(parts,f"{id_}.mp4")
            if os.path.exists(p):
                d=float(subprocess.run(["ffprobe","-v","error","-show_entries","format=duration","-of","csv=p=0",p],capture_output=True,text=True).stdout)
                with tempfile.TemporaryDirectory() as td:
                    for k,ss in (("head",0),("tail",max(0,d-2.5))):
                        wav=f"{td}/{k}.wav"; subprocess.run(["ffmpeg","-v","error","-y","-ss",f"{ss}","-t","2.5","-i",p,"-vn","-ac","1","-ar","16000",wav],check=True)
                        subprocess.run(["mlx_whisper",wav,"--model","mlx-community/whisper-small.en-mlx","--output-format","txt","--output-dir",td,"--output-name",k,"--verbose","False"],capture_output=True)
                        heard=norm(open(f"{td}/{k}.txt").read())
                        want=norm(first[2]) if k=="head" else norm(last[2])
                        if want[:4] not in heard: flags.append(f"{k.upper()} of rendered part does not contain '{first[2] if k=='head' else last[2]}' (heard: {heard[:40]})")
        if flags: nflag+=1; print(f"{id_:5} {ci:8.2f}–{co:8.2f}  " + " | ".join(flags))
        fixed.append([sec,id_,round(na,2),round(nb,2)])
    print(f"{nflag} of {len(cuts)} cuts flagged")
    if fixout: json.dump(fixed,open(fixout,"w")); print("wrote",fixout)
if __name__=="__main__": main()