ffmpeg-skill 1.15.1 → 1.16.1

package/scripts/_common/probe.py

@@ -10,9 +10,9 @@ import math
 import os
 import re
 from fractions import Fraction
-from typing import Any, Dict, List, Optional, Sequence
+from typing import Any, Dict, List, Optional, Sequence, Tuple
 from _common.emit import die
-from _common.runner import STATE, require_tool, run, run_analysis
+from _common.runner import STATE, dry_run_input_pending, require_tool, run, run_analysis
 
 
 def fingerprint(path: str) -> Dict[str, Any]:
@@ -380,3 +380,93 @@ def _aspect_string(w: Optional[int], h: Optional[int]) -> Optional[str]:
         return None
     f = Fraction(w, h)
     return f"{f.numerator}:{f.denominator}"
+
+
+# --------------------------------------------------------------- structure detectors (1.16)
+# silencedetect and scdet, lifted out of silence.py and scenes.py byte-for-byte in 1.16.0 so that
+# metadata.py --auto-chapters can measure structure without importing another tool (no script in
+# scripts/ imports a sibling tool; only the _-prefixed modules are shared). silence.py and
+# scenes.py import them back from here, so their behaviour is unchanged.
+
+SIL_RE = re.compile(r"silence_(start|end): ([0-9.]+)")
+
+
+def detect_silences(path: str, threshold: float, min_silence: float) -> "List[Tuple[float, float]]":
+    if dry_run_input_pending(path):
+        return []
+    ffmpeg = require_tool("ffmpeg")
+    cmd = [ffmpeg, "-hide_banner", "-nostdin", "-i", path, "-vn", "-af",
+           f"silencedetect=noise={threshold}dB:d={min_silence}", "-f", "null", "-"]
+    proc = run_analysis(cmd, check=False, record=True)
+    if proc.returncode != 0:
+        die(f"silencedetect failed:\n{proc.stderr.strip()[-800:]}", kind="ffmpeg")
+    silences: "List[Tuple[float, float]]" = []
+    start = None
+    for kind, val in SIL_RE.findall(proc.stderr):
+        if kind == "start":
+            start = float(val)
+        elif start is not None:
+            silences.append((start, float(val)))
+            start = None
+    if start is not None:  # silence runs to the end
+        silences.append((start, float("inf")))
+    return silences
+
+
+SCORE_RE = re.compile(r"frame:(\d+)\s+pts:\d+\s+pts_time:([0-9.]+)")
+
+
+def detect_scenes(path: str, threshold: float, min_len: float, duration: float, ratio: float = 3.0) -> "List[float]":
+    """Scene cuts = frames whose scdet score is above `threshold` AND stands out from its
+    neighbourhood (score > ratio x median of the surrounding +-12 frames). Sustained motion,
+    flashes and fast pans raise the score on many consecutive frames and are rejected;
+    a real cut is a one-frame spike. On real footage this roughly doubles precision at
+    equal recall compared with the raw scdet threshold."""
+    ffmpeg = require_tool("ffmpeg")
+    proc = run_analysis([ffmpeg, "-hide_banner", "-nostdin", "-i", path, "-an", "-vf",
+                         "scale=320:-2,scdet=threshold=0,metadata=print:file=-", "-f", "null", "-"])
+    # No `sc_pass=1` on scdet: on FFmpeg 5.x that option means "pass only the frames whose
+    # score exceeds the threshold", so every truly static frame (score exactly 0 -- a title
+    # card, colour bars) is dropped before metadata=print and the frame numbers are re-counted
+    # without them. The +-12-frame neighbourhood around a real cut then fills with the moving
+    # segment's scores instead of the still one's zeros, the cut fails the ratio test, and a
+    # 4 s smptebars scene made the cuts on both sides of it disappear (found by the 5.1.1 CI
+    # job, #146). 6.1+ passes every frame either way. Scores are still indexed by frame number
+    # and any frame the filter did not report counts as 0, so a build that drops frames again
+    # cannot shift the neighbourhood.
+    by_frame: "Dict[int, Tuple[float, float]]" = {}
+    cur = None
+    for line in proc.stdout.splitlines():
+        m = SCORE_RE.match(line)
+        if m:
+            cur = (int(m.group(1)), float(m.group(2)))
+            continue
+        if line.startswith("lavfi.scd.score=") and cur is not None:
+            try:
+                by_frame[cur[0]] = (cur[1], float(line.split("=", 1)[1]))
+            except ValueError:
+                pass
+    cuts = [0.0]
+    if not by_frame:
+        return cuts
+    n_frames = max(by_frame) + 1
+    times: "List[float]" = [by_frame[i][0] if i in by_frame else -1.0 for i in range(n_frames)]
+    scores: "List[float]" = [by_frame[i][1] if i in by_frame else 0.0 for i in range(n_frames)]
+    w = 12
+    for i, sc in enumerate(scores):
+        if sc < threshold:
+            continue
+        lo, hi = max(0, i - w), min(len(scores), i + w + 1)
+        neigh = sorted(scores[lo:i] + scores[i + 1:hi])
+        med = neigh[len(neigh) // 2] if neigh else 0.0
+        if sc < ratio * max(med, 0.5):
+            continue
+        # keep only the local maximum inside +-2 frames
+        if any(scores[j] > sc for j in range(max(0, i - 2), min(len(scores), i + 3)) if j != i):
+            continue
+        t = times[i]
+        if t - cuts[-1] >= min_len:
+            cuts.append(t)
+    if duration - cuts[-1] < min_len and len(cuts) > 1:
+        cuts.pop()
+    return cuts

package/scripts/_common/text.py

@@ -440,7 +440,14 @@ ADVANCE_EM = {"ja": 1.0, "zh": 1.0, "ko": 1.0, "th": 1.0, "hi": 0.7, "ar": 0.6,
 
 
 # Scripts written without spaces: a line breaks between any two characters.
-NO_SPACE_SCRIPTS = ("ja", "zh", "ko", "th")
+# Scripts a line may break inside a run of, one character at a time. Thai is deliberately NOT
+# here since 1.16.1: it writes no space inside a phrase, and without a dictionary the wrapper
+# cannot see where one word ends -- every character-level break it took in eval 17 landed inside
+# a word. A Thai run is therefore one atom, broken only at the spaces (or the manual `|`) the
+# writer put there; an over-long run stays long on its own line, the rule long Latin words
+# already follow.
+NO_SPACE_SCRIPTS = ("ja", "zh", "ko")
+NO_BOUNDARY_SCRIPTS = ("th",)   # per-character breaking would chop words: keep the run whole
 
 
 # Per-character Latin advances in em, read off DejaVu Sans (the default caption family, and close
@@ -978,3 +985,547 @@ def drawtext_text_opts(text: str, tmpdir: "Optional[str]" = None) -> str:
     path = os.path.join(tmpdir, name)
     _DRAWTEXT_PENDING[path] = cleaned
     return f"textfile={escape_filter_path(path)}:expansion=none"
+
+
+# --------------------------------------------------------------- caption line breaking (1.16)
+# Lifted out of caption.py in 1.16.0 so graphics.py can wrap the same way (caption.py keeps the
+# names it exported, re-imported from here). The whole breaker is pure: a string in, a list of
+# lines out, no subprocess and no probe, which is what makes the eval regression corpus cheap
+# to lock down in unit tests.
+
+# How much of the frame width a caption line may use. libass's own default SRT margins are 10 of a
+# 384-wide script (2.6 % a side); 5 % a side is the safe area every platform check in this repo uses.
+SAFE_WIDTH_FRACTION = 0.9
+# ORPHAN_MIN_EM: one full-width CJK/Thai character plus a hair. A last line narrower than this is a
+# single stranded character -- eval 14's th1 (a lone 'ล') and dl3 (a lone '行').
+ORPHAN_MIN_EM = 1.1
+
+WRAP_MODES = ("phrase", "measured")
+
+# R3's Japanese preference table. These are *preferences* applied only among positions that
+# already fit the line, so the table can never make a line too wide or change the line count.
+#
+# JA_PARTICLES is a "do not strand at the start of a line" table, which is the direction kinsoku
+# practice actually goes: a particle is enclitic -- it attaches to the word BEFORE it and marks
+# that word's role -- so a line beginning with は or が reads as a fragment torn off its phrase.
+# A break AFTER a particle is therefore preferred (the particle stays with what it marks) and a
+# break BEFORE one is forbidden. The list is the eight case/topic particles named in the 1.16.0
+# task brief (は が を に で と の へ) plus も や から まで より, which a reader of Japanese would
+# add for the same reason. It is a judgement call with no upstream source; treat it as tunable
+# data, not as grammar.
+JA_PARTICLES = "はがをにでとのへもや"              # a break AFTER one of these is preferred, BEFORE one forbidden
+# The multi-character members of the same table. They are matched as whole strings against the
+# text on each side of a candidate break -- putting them in the character string above turned
+# か, ら, ま, で, よ and り into one-character particles of their own, which none of them is.
+JA_PARTICLE_WORDS = ("から", "まで", "より")
+JA_SENTENCE_END = "。、！？」』）"                  # a break AFTER one of these is preferred
+# Characters that may never start a line: small kana, the prolonged sound mark, closing brackets
+# and the Japanese punctuation that hangs on the end of the line before it.
+JA_NO_LINE_START = "ぁぃぅぇぉっゃゅょァィゥェォッャュョーヽヾゝゞ、。！？）」』】〕》’”％"
+JA_NO_LINE_END = "（「『【〔《‘“"                   # ... and the ones that may never end a line
+
+# R4. Function words belong to the phrase that FOLLOWS them: an article or preposition begins the
+# noun phrase it governs, so a break before one is the good break (the word opens the next line
+# with its phrase) and a break after one is the bad break (it is stranded at the end of a line,
+# away from what it governs). Both directions are scored, which is what makes the rule decide
+# rather than merely veto. Frozen data, matched case-folded on the atom with its punctuation
+# stripped; six languages because those are the Latin-script languages the eval corpus covers. A
+# word in several sets means the same thing structurally in each, so the union is used when no
+# --lang was given.
+FUNCTION_WORDS = {
+    "en": {"a", "an", "the", "of", "to", "in", "on", "at", "for", "with", "by", "from", "and",
+           "or", "as", "is", "it", "its", "this", "that", "into", "than", "but", "so"},
+    "es": {"el", "la", "los", "las", "un", "una", "unos", "unas", "de", "del", "al", "en", "con",
+           "por", "para", "y", "o", "que", "su", "sus", "lo", "se", "es"},
+    "pt": {"o", "a", "os", "as", "um", "uma", "de", "do", "da", "dos", "das", "em", "no", "na",
+           "nos", "nas", "com", "por", "para", "e", "que", "se", "ao", "aos"},
+    "fr": {"le", "la", "les", "un", "une", "de", "du", "des", "à", "au", "aux", "en", "dans",
+           "et", "ou", "que", "qui", "ce", "ces", "son", "sa", "ses", "par", "pour", "avec", "sur"},
+    "de": {"der", "die", "das", "ein", "eine", "einen", "einem", "einer", "den", "dem", "des",
+           "zu", "in", "im", "auf", "mit", "und", "oder", "von", "vom", "für", "aus", "an"},
+    "it": {"il", "lo", "la", "i", "gli", "le", "un", "una", "uno", "di", "del", "della", "da",
+           "in", "nel", "con", "per", "e", "che", "su", "al", "ai"},
+}
+_FUNCTION_WORDS_ANY = frozenset().union(*FUNCTION_WORDS.values())
+
+# Penalty scores. Only the ordering matters; 1.0 means "never choose this if anything else fits".
+PENALTY_FORBIDDEN = 1.0
+PENALTY_OKURIGANA = 0.9       # between a kanji stem and the hiragana that inflects it
+PENALTY_FUNCTION_WORD = 0.8   # R4: the line before the break ends in an article/preposition
+PENALTY_IDEOGRAPHS = 0.6      # between two kanji: no evidence either way, mildly discouraged
+PENALTY_NEUTRAL = 0.5         # between two content words, or two characters with nothing to say
+PENALTY_FUNCTION_WORD_START = 0.2  # R4: the next line opens with the article/preposition it governs
+PENALTY_PARTICLE = 0.2        # R3: after a particle, so the particle stays with the word it marks
+PENALTY_SENTENCE_END = 0.0    # R3: after 。、！？ -- the one break a reader expects
+
+_HYPHENS = ("-", "‐")    # ‑ (non-breaking hyphen) is deliberately NOT here
+
+
+def _atoms(line: str) -> "List[Tuple[str, bool]]":
+    """Break a line into the smallest pieces a wrap may separate -- one atom per CJK/Thai
+    character, one per emoji cluster, one per whitespace-delimited word otherwise -- each with
+    whether a space stood before it in the original. The flag is what puts the text back together
+    exactly as written: "Hello 世界" keeps its space, "世界です" gains none."""
+    out: "List[Tuple[str, bool]]" = []
+    word = ""
+    spaced = False        # a space stands before the atom being built
+    pending = False       # a space stands before the NEXT atom
+    attach_next = False   # a leading Thai/Lao vowel is waiting for its base consonant
+    # An emoji cluster is one atom: a wrap must never land inside a ZWJ sequence, a flag pair or
+    # between a base and its skin-tone modifier (the same rule combining marks already follow).
+    clusters = {i: len(cl) for i, cl in emoji_clusters(line)}
+    i = 0
+    while i < len(line):
+        ch = line[i]
+        if i in clusters:
+            cluster = line[i:i + clusters[i]]
+            if word:
+                out.append((word, spaced))
+                word = ""
+            out.append((cluster, pending))
+            pending = False
+            attach_next = False
+            i += clusters[i]
+            continue
+        i += 1
+        if char_script(ch) in NO_SPACE_SCRIPTS:
+            if word:
+                out.append((word, spaced))
+                word = ""
+            if out and not pending and _is_katakana_run(ch) and _is_katakana_run(out[-1][0][-1]):
+                # a katakana word (タイミング, コンピューター) is one atom: eval 17 saw タイ|ミング
+                out[-1] = (out[-1][0] + ch, out[-1][1])
+            elif out and (attach_next or _is_mark(ch)):
+                # never break between a base and the mark (or the leading vowel) that belongs to
+                # it: the line would start with an orphaned tone mark or vowel sign
+                out[-1] = (out[-1][0] + ch, out[-1][1])
+            else:
+                out.append((ch, pending))
+                pending = False
+            attach_next = ord(ch) in LEADING_VOWELS
+        elif ch.isspace():
+            if word:
+                out.append((word, spaced))
+                word = ""
+            pending = True
+        else:
+            if not word:
+                spaced, pending = pending, False
+            word += ch
+    if word:
+        out.append((word, spaced))
+    return out
+
+
+def _split_hyphens(atoms: "List[Tuple[str, bool]]") -> "List[Tuple[str, bool]]":
+    """R1's one addition to the atom list: a hyphenated token may break *after* its hyphen.
+
+    "end-to-end" becomes `end-` / `to-` / `end`, each piece carrying the space flag of the token
+    it came from for the first piece and False for the rest, so _join() puts it back with no space
+    at all. A hyphen that is the first or last character of the token (`-5`, `well-`) is never a
+    break point: the guard is that both sides must be non-empty."""
+    out: "List[Tuple[str, bool]]" = []
+    for atom, spaced in atoms:
+        if len(atom) < 3 or not any(h in atom[1:-1] for h in _HYPHENS):
+            out.append((atom, spaced))
+            continue
+        piece = ""
+        first = True
+        for i, ch in enumerate(atom):
+            piece += ch
+            if ch in _HYPHENS and 0 < i < len(atom) - 1:
+                out.append((piece, spaced if first else False))
+                piece = ""
+                first = False
+        if piece:
+            out.append((piece, spaced if first else False))
+    return out
+
+
+def _join(left: str, atom: str, spaced: bool) -> str:
+    """Put an atom back on a line, restoring the space that stood before it."""
+    if not left:
+        return atom
+    return left + (" " if spaced else "") + atom
+
+
+def _break_spaced(first: str, second: str) -> bool:
+    """Did a space stand at the break between these two wrapped lines? Only spaced scripts put one
+    there -- a CJK/Thai break sits between two characters that were written with nothing between
+    them, and re-joining them with a space would insert a character the cue never had."""
+    if not first or not second:
+        return False
+    return char_script(first[-1]) not in NO_SPACE_SCRIPTS and char_script(second[0]) not in NO_SPACE_SCRIPTS \
+        and char_script(first[-1]) != "emoji" and char_script(second[0]) != "emoji"
+
+
+def _is_kana(ch: str) -> bool:
+    return 0x3040 <= ord(ch) <= 0x30FF
+
+
+def _is_katakana_run(ch: str) -> bool:
+    """Katakana proper plus the prolonged-sound mark: the characters one loan word is made of."""
+    cp = ord(ch)
+    return (0x30A1 <= cp <= 0x30FA) or cp == 0x30FC or (0x31F0 <= cp <= 0x31FF) or (0xFF66 <= cp <= 0xFF9F)
+
+
+def _is_hiragana(ch: str) -> bool:
+    return 0x3040 <= ord(ch) <= 0x309F
+
+
+def _is_ideograph(ch: str) -> bool:
+    cp = ord(ch)
+    return 0x3400 <= cp <= 0x4DBF or 0x4E00 <= cp <= 0x9FFF or 0xF900 <= cp <= 0xFAFF
+
+
+def _is_weak_line(line: str) -> "bool":
+    """A line no reader should be given on its own (R2).
+
+    1.15 asked only "is the last line one atom narrower than ORPHAN_MIN_EM", which a full-width
+    character passes: dl3 still showed a lone `2` and a stranded `行`. Three cases instead, any of
+    which makes a line too thin to stand alone:
+      - a single character narrower than ORPHAN_MIN_EM (1.15's rule, kept);
+      - nothing but digits, punctuation and symbols, at most two characters ("2", "--");
+      - a single kana, whatever its width -- a kana is a full em and passes the width test, but a
+        line holding one is a syllable, not a word.
+    """
+    stripped = (line or "").strip()
+    if not stripped:
+        return True
+    if len(stripped) == 1 and text_width_em(stripped) < ORPHAN_MIN_EM:
+        return True
+    if len(stripped) <= 2 and all(unicodedata.category(c)[0] in "NPS" for c in stripped):
+        return True
+    if len(stripped) == 1 and char_script(stripped) == "ja" and _is_kana(stripped):
+        return True
+    return False
+
+
+def _function_words(lang: "Optional[str]") -> "frozenset":
+    """R4's table for this language. An unknown or absent language gets the union of the six sets:
+    a token that appears in several of them is the same kind of word in each, which is why the
+    rule is a penalty and not a refusal."""
+    key = (lang or "").strip().lower().split("-")[0]
+    if key in FUNCTION_WORDS:
+        return frozenset(FUNCTION_WORDS[key])
+    return _FUNCTION_WORDS_ANY
+
+
+def _bare_word(atom: str) -> str:
+    return "".join(c for c in (atom or "") if c.isalpha() or c == "'").strip("'").lower()
+
+
+def _particle_starts(text: str) -> bool:
+    """Does `text` begin with a particle -- one character, or one of the two-character ones?"""
+    if not text:
+        return False
+    return text[0] in JA_PARTICLES or text.startswith(JA_PARTICLE_WORDS)
+
+
+def _particle_ends(text: str) -> bool:
+    """Does `text` end with a particle? `から` counts, a bare `ら` does not."""
+    if not text:
+        return False
+    return text[-1] in JA_PARTICLES or text.endswith(JA_PARTICLE_WORDS)
+
+
+def break_penalty(prev_char: str, next_char: str, lang: "Optional[str]" = None,
+                  before: str = "", after: str = "") -> float:
+    """How bad a break between these two characters is, 0.0 (preferred) to 1.0 (forbidden).
+
+    Only consulted among break positions that already fit `max_em`, so a preference can never
+    widen a line or change the line count. Japanese gets the particle half of the table -- a break
+    AFTER a particle is preferred and a break BEFORE one forbidden, because a particle attaches to
+    the word before it; Chinese gets only the sentence-end and forbidden halves, because particles
+    are Japanese grammar.
+
+    `before`/`after` are the text on each side of the break when the caller has it, which is what
+    lets the two-character particles (から/まで/より) be matched as words. Without them only the
+    single-character table applies."""
+    if not prev_char or not next_char:
+        return PENALTY_NEUTRAL
+    script = (lang or "").strip().lower().split("-")[0]
+    if script not in ("ja", "zh"):
+        # A kana on either side settles it: only Japanese has them, and char_script() reads a bare
+        # Han character as Chinese, which used to switch the particle rules off for exactly the
+        # break they exist to judge (`...が|決まる` -- kana before, kanji after).
+        if _is_kana(prev_char) or _is_kana(next_char):
+            script = "ja"
+        else:
+            script = char_script(next_char)
+            if script not in ("ja", "zh"):
+                script = char_script(prev_char)
+    if next_char in JA_NO_LINE_START or prev_char in JA_NO_LINE_END or _is_mark(next_char):
+        return PENALTY_FORBIDDEN
+    if script not in ("ja", "zh"):
+        return PENALTY_NEUTRAL
+    if prev_char in JA_SENTENCE_END:
+        return PENALTY_SENTENCE_END
+    if script == "ja" and _particle_starts(after or next_char):
+        # a particle may not open a line: it belongs to the word before it (kinsoku)
+        return PENALTY_FORBIDDEN
+    if script == "ja" and _particle_ends(before or prev_char):
+        return PENALTY_PARTICLE
+    if script == "ja" and _is_ideograph(prev_char) and _is_hiragana(next_char):
+        # okurigana: 決|まる is inside a word even though neither half is a "word" on its own
+        return PENALTY_OKURIGANA
+    if _is_ideograph(prev_char) and _is_ideograph(next_char):
+        return PENALTY_IDEOGRAPHS
+    return PENALTY_NEUTRAL
+
+
+def _cut_penalty(atoms: "Sequence[Tuple[str, bool]]", cut: int, lang: "Optional[str]") -> float:
+    """The penalty of breaking `atoms` before index `cut`."""
+    prev_atom = atoms[cut - 1][0]
+    next_atom = atoms[cut][0]
+    if not prev_atom or not next_atom:
+        return PENALTY_NEUTRAL
+    if atoms[cut][1]:
+        # a space stood here: a spaced script, so R4 is the rule that applies, in both directions
+        if all(not ch.isalnum() for ch in next_atom):
+            return PENALTY_FORBIDDEN   # never strand punctuation at the start of a line
+        words = _function_words(lang)
+        if _bare_word(prev_atom) in words:
+            return PENALTY_FUNCTION_WORD        # stranded at the end of a line, away from its noun
+        if _bare_word(next_atom) in words:
+            return PENALTY_FUNCTION_WORD_START  # opens the next line with the phrase it governs
+        return PENALTY_NEUTRAL
+    if prev_atom.endswith(_HYPHENS):
+        return PENALTY_NEUTRAL         # R1: a hyphen is a legitimate break point
+    # the text on each side, so a two-character particle (から/まで/より) is seen as one
+    before = "".join(a for a, _sp in atoms[:cut])
+    after = "".join(a for a, _sp in atoms[cut:])
+    return break_penalty(prev_atom[-1], next_atom[0], lang, before=before, after=after)
+
+
+def best_break(atoms: "Sequence[Tuple[str, bool]]", max_em: float,
+               lang: "Optional[str]" = None) -> "Optional[int]":
+    """The index to break `atoms` at so they become two lines, or None when none fits.
+
+    Among every position whose two halves both fit `max_em`, the one minimising
+    (penalty, widest line, |width difference|) wins: R1-R4 choose first, and 1.15's
+    minimise-the-widest-line rule breaks the ties it used to decide alone."""
+    best = None
+    for cut in range(1, len(atoms)):
+        a = b = ""
+        for atom, sp in atoms[:cut]:
+            a = _join(a, atom, sp)
+        for atom, sp in atoms[cut:]:
+            b = _join(b, atom, sp)
+        wa, wb = text_width_em(a), text_width_em(b)
+        if max(wa, wb) > max_em:
+            continue
+        if _is_weak_line(a) or _is_weak_line(b):
+            continue
+        key = (_cut_penalty(atoms, cut, lang), max(wa, wb), abs(wa - wb))
+        if best is None or key < best[0]:
+            best = (key, cut)
+    return None if best is None else best[1]
+
+
+def _fix_orphans(lines: "List[str]", max_em: float) -> "List[str]":
+    """No last line that is a single stranded atom.
+
+    Greedy wrapping leaves one character alone whenever the line before it filled exactly: eval 14
+    produced a Thai cue ending in a lone `ล` and a Japanese one ending in a lone `行`. While the
+    last line is one atom narrower than ORPHAN_MIN_EM, the last atom of the line above moves down
+    onto it -- but only while the result still fits and the line above does not become an orphan
+    itself, so a two-word cue is never made worse."""
+    lines = list(lines)
+    for _ in range(len(lines)):
+        if len(lines) < 2:
+            break
+        tail = _atoms(lines[-1])
+        if len(tail) != 1 or text_width_em(lines[-1]) >= ORPHAN_MIN_EM:
+            break
+        prev = _atoms(lines[-2])
+        if len(prev) < 2:
+            break
+        moved, spaced = prev[-1]
+        new_prev = ""
+        for atom, sp in prev[:-1]:
+            new_prev = _join(new_prev, atom, sp)
+        new_last = _join(moved, tail[0][0], _break_spaced(lines[-2], lines[-1]))
+        if text_width_em(new_last) > max_em or text_width_em(new_prev) < ORPHAN_MIN_EM:
+            break
+        lines[-2], lines[-1] = new_prev, new_last
+    return lines
+
+
+def _fix_weak_lines(lines: "List[str]", max_em: float) -> "List[str]":
+    """R2, generalised: _fix_orphans run at *every* boundary, against _is_weak_line.
+
+    1.15 only ever looked at the last line, so a stranded digit or kana in the middle of a
+    three-line cue survived. Walking upward from the last line, while a line is weak the last atom
+    of the line above moves down onto it -- with 1.15's two guards intact (the result must still
+    fit, and the line above must not itself become weak), so the line count never changes."""
+    lines = list(lines)
+    for i in range(len(lines) - 1, 0, -1):
+        for _ in range(len(lines)):
+            if not _is_weak_line(lines[i]):
+                break
+            prev = _atoms(lines[i - 1])
+            if len(prev) < 2:
+                break
+            moved, _spaced = prev[-1]
+            new_prev = ""
+            for atom, sp in prev[:-1]:
+                new_prev = _join(new_prev, atom, sp)
+            new_last = _join(moved, lines[i], _break_spaced(lines[i - 1], lines[i]))
+            if text_width_em(new_last) > max_em or _is_weak_line(new_prev) \
+                    or text_width_em(new_prev) < ORPHAN_MIN_EM:
+                break
+            lines[i - 1], lines[i] = new_prev, new_last
+    return lines
+
+
+def _rebalance(lines: "List[str]", max_em: float) -> "List[str]":
+    """Move each break to the one that minimises the widest line of the pair, without changing the
+    line count.
+
+    Greedy wrapping fills line 1 to the brim and leaves line 2 short, which is what split eval 14's
+    `"A third line the tool times for me"` mid-phrase. Only spaced scripts are rebalanced: a
+    non-spaced script has no phrase structure in its atom list, so moving the break there only
+    moves the ragged edge. A break is never placed before a punctuation-only atom."""
+    if len(lines) < 2:
+        return lines
+    out = list(lines)
+    for i in range(len(out) - 1):
+        first, second = out[i], out[i + 1]
+        tail_atoms = _atoms(second)
+        if tail_atoms:
+            tail_atoms[0] = (tail_atoms[0][0], _break_spaced(first, second))
+        atoms = _atoms(first) + tail_atoms
+        if not atoms or any(char_script(ch) in NO_SPACE_SCRIPTS for ch in first + second):
+            continue
+        best = None
+        for cut in range(1, len(atoms)):
+            if not atoms[cut][1]:
+                continue  # only break where a space stood
+            if all(not ch.isalnum() for ch in atoms[cut][0]):
+                continue  # never strand punctuation at the start of a line
+            a = b = ""
+            for atom, sp in atoms[:cut]:
+                a = _join(a, atom, sp)
+            for atom, sp in atoms[cut:]:
+                b = _join(b, atom, sp)
+            wa, wb = text_width_em(a), text_width_em(b)
+            if max(wa, wb) > max_em:
+                continue
+            key = (max(wa, wb), abs(wa - wb))
+            if best is None or key < best[0]:
+                best = (key, a, b)
+        if best is not None:
+            out[i], out[i + 1] = best[1], best[2]
+    return out
+
+
+def _rebalance_phrase(lines: "List[str]", max_em: float, lang: "Optional[str]") -> "Tuple[List[str], int]":
+    """_rebalance with R1-R4 deciding, for every script rather than spaced ones only.
+
+    Returns the new lines and how many breaks a phrase rule moved away from the position 1.15's
+    widest-line rule alone would have chosen -- the `phrase_breaks` count in the result."""
+    if len(lines) < 2:
+        return list(lines), 0
+    out = list(lines)
+    moved = 0
+    for i in range(len(out) - 1):
+        first, second = out[i], out[i + 1]
+        tail_atoms = _atoms(second)
+        if tail_atoms:
+            tail_atoms[0] = (tail_atoms[0][0], _break_spaced(first, second))
+        atoms = _split_hyphens(_atoms(first) + tail_atoms)
+        if len(atoms) < 2:
+            continue
+        cut = best_break(atoms, max_em, lang)
+        if cut is None:
+            continue
+        a = b = ""
+        for atom, sp in atoms[:cut]:
+            a = _join(a, atom, sp)
+        for atom, sp in atoms[cut:]:
+            b = _join(b, atom, sp)
+        if (a, b) != (first, second):
+            moved += 1
+        out[i], out[i + 1] = a, b
+    return out, moved
+
+
+def _greedy_chunks(raw: str, max_em: float) -> "List[str]":
+    """The greedy fill on its own: the line count every mode must keep."""
+    current = ""
+    chunk: "List[str]" = []
+    for atom, spaced in _atoms(raw):
+        candidate = _join(current, atom, spaced)
+        if current and text_width_em(candidate) > max_em:
+            chunk.append(current)
+            current = atom
+        else:
+            current = candidate
+    if current:
+        chunk.append(current)
+    return chunk
+
+
+def _balance(chunk: "List[str]", max_em: float, mode: str, lang: "Optional[str]") -> "List[str]":
+    """The post-passes for one greedy chunk, in the mode's own order. Never changes the count:
+    a pass that would is discarded, exactly as 1.15 did."""
+    if len(chunk) < 2:
+        return chunk
+    if mode == "measured":
+        fixed = _fix_orphans(chunk, max_em)
+        rebalanced = _rebalance(fixed, max_em)
+    else:
+        fixed = _fix_weak_lines(_fix_orphans(chunk, max_em), max_em)
+        rebalanced, _moved = _rebalance_phrase(fixed, max_em, lang)
+        rebalanced = _fix_weak_lines(rebalanced, max_em)
+    if len(rebalanced) == len(chunk):
+        return rebalanced
+    return fixed if len(fixed) == len(chunk) else chunk
+
+
+def wrap_text(text: str, max_em: float, *, balance: bool = True, mode: str = "phrase",
+              lang: "Optional[str]" = None) -> "List[str]":
+    """Wrap `text` to lines no wider than `max_em` em, keeping the manual breaks it already has.
+
+    An atom wider than the whole line (one very long word) is left alone on its line rather than
+    cut mid-word: an over-long line is readable, a chopped word is not.
+
+    `mode="phrase"` (the default since 1.16) then applies the four phrase rules -- never inside a
+    word or across a hyphen's wrong side (R1), no line that is a lone digit, punctuation or kana
+    (R2), Japanese/Chinese breaks preferred at sentence ends and after particles, never before one
+    and never inside a word (R3), and an article or preposition kept with the phrase it governs by
+    preferring the break before it and avoiding the break after it (R4). `mode="measured"` is
+    1.15's behaviour exactly: no one-character orphan line, and a break chosen only to minimise the
+    widest line. Neither mode ever changes the number of lines the greedy fill produced.
+    """
+    lines: "List[str]" = []
+    for raw in text.split("\n"):
+        if not raw.strip():
+            continue
+        chunk = _greedy_chunks(raw, max_em)
+        lines.extend(_balance(chunk, max_em, mode, lang) if balance else chunk)
+    return lines or [text]
+def wrap_variants(text: str, max_em: float, *, mode: str = "phrase",
+                  lang: "Optional[str]" = None) -> "Tuple[List[str], List[str], List[str]]":
+    """`(wrapped, greedy, measured)` for one cue from a single greedy fill.
+
+    layout_cues needs all three -- `wrapped` is what is burnt in, `greedy` is what `rebalanced`
+    counts against and `measured` what `phrase_breaks` counts against -- and used to call
+    wrap_text() three times, re-running the atomiser and the greedy fill each time. The fill is
+    the same for every mode, so it is done once here and only the post-passes are repeated.
+    `measured` is the same list object as `wrapped` when that is already the mode.
+    """
+    wrapped: "List[str]" = []
+    greedy: "List[str]" = []
+    measured: "List[str]" = []
+    for raw in text.split("\n"):
+        if not raw.strip():
+            continue
+        chunk = _greedy_chunks(raw, max_em)
+        greedy.extend(chunk)
+        wrapped.extend(_balance(list(chunk), max_em, mode, lang))
+        measured.extend(chunk if mode == "measured" else _balance(list(chunk), max_em, "measured", None))
+    if not greedy:
+        greedy = [text]
+    return (wrapped or [text], greedy, measured or [text])