`-s` (subtable breaks) crashes on any RTL-only font, because the RTL path prints the LTR subtable counter

[pauldhunt]

Summary

The bug, in plain terms. The kern writer counts subtables so it can print N subtables created. It keeps two separate tallies — one for left-to-right kerning, one for right-to-left. It increments the correct one, then prints the left-to-right one no matter which direction it just wrote.

Why that's fatal. Each tally is initialised inside its own direction's branch. A font with no left-to-right pairs never enters the left-to-right
branch, so that tally is never created, and the print statement asks for a number that doesn't exist:

AttributeError: 'run' object has no attribute 'num_subtables'

It fails before the output file is opened, so -s writes no kern.fea at all. Not a cosmetic bug.

The quieter half. In a font that has both directions, nothing crashes, but the right-to-left runs print the left-to-right number — so a font writing two right-to-left subtables cheerfully reports 0 subtables created, twice.

Why it matters. -s writes subtable breaks, one of the standard ways to keep a large kern feature inside GPOS's 16-bit subtable offset limit. With -s unusable, useExtension is the only remedy left for an RTL-only font.

Reproducer and a verified one-line fix below. Related: #26, which reports the same print for a different reason.

Full report:

kernFeatureWriter.py — _build_st_output(), defined line 1126. All line numbers below are upstream at 503b206, so they can be quoted into the issue unchanged; the vendored copy at scripts/kern/kernFeatureWriter.py is offset by its banner (+19 lines at the time of writing, and the banner is edited from time to time, so recompute rather than trust that number).

_build_st_output() keeps two counters, one per direction, and increments the right one:

if rtl:
    self.num_subtables_rtl += 1
    ...
else:
    self.num_subtables += 1
    ...
print(f'{self.num_subtables} subtables created')      # line 1148 — LTR counter, both paths

Both counters are initialised late and separately, each inside its own direction's branch of _make_fea_data():

if any(ltr_container_dicts):
    ...
    if self.write_subtables:
        self.num_subtables = 0                        # line 1221 — LTR branch only
...
if any(rtl_container_dicts):
    ...
    if self.write_subtables:
        self.num_subtables_rtl = 0                    # line 1256 — RTL branch only

A font with no LTR pairs never enters the first branch, so self.num_subtables is never created. The RTL call to _build_st_output() then reaches line 1148 and raises:

AttributeError: 'run' object has no attribute 'num_subtables'

Nothing is written — the exception fires inside _make_fea_data(), before the output file is opened, so -s produces no kern.fea at all on such a font.

The same line is wrong in the non-crashing case. _build_st_output() runs four times per write — LTR glyph-to-class, LTR class-to-class, then the same two for RTL — and line 1148 prints at the end of every one of them. When a font has both directions, the two RTL calls print the LTR count, so a font whose subtables are all RTL reports 0 subtables created twice while writing two of them.

Read the counter as a running total, not a per-call count: it is reset once per direction (lines 1221 / 1256) and incremented once per non-empty subtable dict, so consecutive calls print 1, then 2, and the subtable; break is emitted only from the second one onward.

Reproduction

import defcon

f = defcon.Font()
for n in ('alef', 'beh', 'teh', 'reh'):
    f.newGlyph(n).width = 500

f.groups['public.kern1.ALEF_ARA'] = ['alef']          # _ARA suffix marks the pair RTL
f.groups['public.kern2.BEH_ARA'] = ['beh', 'teh']

f.kerning[('public.kern1.ALEF_ARA', 'public.kern2.BEH_ARA')] = -50
f.kerning[('reh', 'public.kern2.BEH_ARA')] = -30
f.save('ReproRTL.ufo')

python kernFeatureWriter.py -s ReproRTL.ufo writes no kern.fea and exits 1 with:

  File ".../kernFeatureWriter.py", line 1148, in _build_st_output
    print(f'{self.num_subtables} subtables created')
AttributeError: 'run' object has no attribute 'num_subtables'

The LTR control needs glyphs, not just a pair. Adding f.kerning[('A', 'V')] = -40 to the fixture above changes nothing — A and V do not exist, KerningSanitizer discards the pair as invalid, ltr_container_dicts stays empty, and the run crashes exactly as before. Create the two glyphs as well:

for n in ('A', 'V'):
    f.newGlyph(n).width = 500
f.kerning[('A', 'V')] = -40

Now the command succeeds, and prints 0 subtables created twice for the two RTL calls — the second half of the bug. The zero is not a sizing artifact: only glyph_group and group_group are routed through MakeMeasuredSubtables (lines 1223–1235 for LTR, 1258–1270 for RTL), while a plain glyph-to-glyph pair such as A V is written directly and never counted, so the LTR counter legitimately stays at 0 and the RTL calls print it.

Without -s the same UFO writes fine, so the fault is in the subtable path only.

Suggested fix

-        print(f'{self.num_subtables} subtables created')
+        count = self.num_subtables_rtl if rtl else self.num_subtables
+        print(f'{count} subtables created')

Verified on the reproducer, 2026-08-26, against upstream 503b206. RTL-only now writes kern.fea — 22 lines, 2 pos rules, one subtable; break — and prints the running RTL total across its two calls:

1 subtables created
2 subtables created

The mixed fixture prints 0, 0, 1, 2, which is correct for it: no LTR subtable exists, and both RTL ones do. Initialising both counters in __init__ instead would stop the crash but keep the wrong count, so the direction-aware read is the better fix; doing both is better still.

Status in this project

Reached, not latent. AdobeUrdu is RTL-only — the CTX regeneration emits one RTL_kerning lookup and no LTR pairs — so kernFeatureWriterCTX.py -s AdobeUrdu-Regular.ufo crashed exactly as above. Subtable breaks are one of the documented ways to keep a large kern feature inside GPOS offset limits (KERN_FEATURE_WRITER_SPEC.md §8, criterion 4), so -s being unusable cost a remedy.

The fix is applied to the vendored scripts/kern/kernFeatureWriter.py as banner delta (2), labelled LOCAL PATCH (UNFILED) and citing this doc where a filed patch cites #N; the filing stays owed in the tracker's Filed column (row 5). Per KERN_FEATURE_WRITER_SPEC.md §6 (the local-patch process) as revised 2026-08-28, the repro doc plus its tracker row is the gate, not the issue number. Do not patch Git/ATOtools/python-modules/kernFeatureWriter.py in place — it is upstream, and the upstream line numbers quoted above describe it as it still stands.

kernFeatureWriterCTX.py -s -o /tmp/kern_s.fea AdobeUrdu-Regular.ufo now exits 0 and writes 2574 lines, 2484 pos rules, one subtable; break, printing 1 subtables created then 2 subtables created for its two RTL calls (2026-08-28). Verified with -o outside the repo so the build's committed kern.fea is not overwritten.