not_supported_extensions.txt

# --- scene.org / Fujiology demoscene-compo oddities (2026-07-10) ---
# These reached the index via URL onboarding tagged with a generic platform
# string ("Demoscene"/"Windows"/"Linux"/"TIC-80") that resolves to no hardware,
# so they surfaced in the "no screenshot / no platform logo" audit. On review
# none has a decoder in our stack -- they are archives, commercial DAW projects,
# executables, or hardware trackers with no open replayer -- so indexing them
# only yields broken GUI entries. Skip them. (mon/ntk from the same audit ARE
# playable and were fixed in MusicDatabase.cpp instead.)

# compression archives, not music formats (compo entries submitted as archives)
.arj
.lzx
.tgz
.xz
.img

# .one -- unidentified container ("modk.one"); no reader
.one

# FL Studio (FruityLoops) project -- commercial DAW, no open renderer (cf. renoise/psy)
.flp

# 32k executable-music intro (Windows PE), not a decodable module
.ps1

# TIC-80 fantasy-console cart -- no TIC-80 music replayer in the stack
.tic

# Dirtywave M8 tracker song -- hardware tracker, no open replayer
.m8s

# unidentified obscure demoscene formats, no decoder anywhere
.legoz
.asg

# .snort -- 2 scene.org "mod.snort" files, no .snort route/decoder
.snort

# (.ac3 Dolby AC-3 and .mpeg raw MPEG audio are NOT skipped -- they are wired
# into FFMPEGPlugin's ffmpeg gate and play as audio.)

# renoise commercial player
.rns
.xrns

# psycle tracker -- no decoder
.psy

# sound club (bluemoon software; PC/DOS+Win, NOT amiga). .sn = DOS 1993,
# .sn2 = win95/nt sequel "sound club 2" (magic SN2\xd2, little-endian IFF-like
# chunks NAM/SEQ/PAT/pna/voi). pattern data is parseable but modules carry NO
# sample/instrument data -- voices come from the closed built-in bluemoon synth,
# so playback needs full RE of the win32 engine. no open replayer anywhere
# (UADE/OpenMPT/libxmp/ZXTune/rePlayer/NostalgicPlayer/vgmstream). OpenMPT's
# "Sound Club 2" string only labels S3M *exports*, it can't read native .sn/.sn2.
.sn
.sn2

# zx spectrum flash tracker -- nothing exists to decode
.fls

# beaver sweeper (modland "Beaver Sweeper", author Steffo, 4 .gtk tunes, magic
# "+SNT"). NOT a live ext-line: .gtk is owned by OpenMPT for Graoumf Tracker
# (magic "GTK"), which plays fine. the +SNT files are an unrelated, undocumented
# format with no open replayer anywhere (OpenMPT/UADE/libxmp/NostalgicPlayer);
# UADE's PTK-Prowiz routes them via the snt prefix but renders pure silence.
# OpenMPTPlugin::canHandle content-gates .gtk on "GTK" so these Skip cleanly
# instead of hard-FAILing. documented here, intentionally commented out.
# .gtk

# .MUS upper case fcatrin/fstr (MIT), is x86 real-mode asm
# targeting AdLib OPL2 (wrong chip, not portable C). 
# commented but documented due to .mus lowercase clash

# axs softsynth/tracker (resolution audio, www.resolutionaudio.nl; closed
# commercial DOS 2.x + Win32 3.x, freeware since 2001). IFF container
# "FORM <size> AXSF" with USER/SHDR/SONG/BLOK/TRAK/INST/INNM chunks. The
# sequencer/pattern data is parseable, but the INST chunks are SYNTH PATCHES
# (a Roland JP8000-style softsynth model) -- there is NO sample/PCM data in the
# file, so voices come entirely from the closed built-in DSP engine. No open
# replayer anywhere (UADE/OpenMPT/libxmp/ZXTune/rePlayer); playback needs full
# RE of the win32/dos synth. Same dead-end as .psy/.sn. ~22 tunes on modland.
.axs

# atari falcon ace tracker -- only falcon or port whole emulator
.am

# shroom tracker -- no decoder
.sho

# jeskola buzz -- modular daw, needs closed win32 x86 machine dlls
.bmx

# piece music driver (pc-98, chemool) -- NOT kaja pmd (which is .m/.m2/.mz);
# distinct custom opna driver, no open replayer exists (98fmplayer, pmdwin,
# mdplayer, rePlayer all reject it). needs full driver RE.
.pmd

# stonetracker (amiga, 8-voice software mixer; .spm=song + .sps=sample bank
# pair, magic SPM\x02 / SPS\x02; modland Stonetracker/). NO open replayer:
# UADE ships no eagleplayer, not in rePlayer/libxmp/OpenMPT. Only source is
# Marty's x86 DOS/GUS hardware-mixing player -- no portable software mixer to
# lift. Needs full playroutine + format RE. Removed from UADE supported_ext so
# they Skip instead of hard-FAIL.
.spm
.sps

# dux (amiga). Was listed in UADE supported_ext but UADE has NO player for it --
# no eagleplayer.conf prefix, no amifilemagic content detection, no player binary
# anywhere in the vendored 3.05 tree. Every .dux module therefore just FAILED
# ("score died"). No open replayer found (UADE/rePlayer/libxmp/OpenMPT none).
# Removed from supported_ext so they Skip instead of hard-FAIL. Re-open if a dux
# eagleplayer lands upstream.
.dux

# 670 (modland "Composer 670 (CDFM)", 7 tunes) -- SHIPPED, no longer unsupported.
# The .670 is the demo/stripped layout of the C67 format: a compact 10-byte
# header (byte0=speed 1-15, orderListLen, numPatterns, numDigInst, numOPLInst,
# loopDest, u32 sampleOffset), no instrument names, and only a pattern *offset*
# table -- so libopenmpt's Load_c67 (which wants the unpacked 1954-byte named
# header + offset/length table) rejects it directly. But the pattern byte-stream
# and the 16-byte PCM / 11-byte OPL instrument structs are byte-identical to C67,
# so OpenMPTPlugin::convert670ToC67 repacks .670 -> C67 in memory (fill names
# with zeros, synthesize the 128-entry length table, pad each pattern to >=3
# bytes, lay patterns in index order so pattern 127 ends exactly at the sample
# blob, append sum(sampleLengths) PCM bytes) and hands the result to libopenmpt.
# canHandle content-gates on a clean conversion. Verified: all 7 modland tunes
# convert and pass Load_c67's ValidateHeader + pattern/sample bounds (3421 note
# events for amnesia-intro, sample region byte-exact). The old "no depacker
# exists" verdict was wrong -- the format is fully documented (shikadi CDFM_Format)
# and Composer's own 670TOC67 converter exists.

# Skale Tracker for PC by Baktery: No open replayer and no public spec exist
.skm

# CyberTracker (C64 SID, CyberBrain/No Name; native magic NNTRKMZX, song PRG @
# $0400). The only public converter (DOS ctpacker BETA) is broken (malloc/DPMI),
# and a native plugin would need full RE of the .ct->SID packer. Low value: the
# composers' own packed SIDs are already in HVSC and play via csidplugin (cSID).
.ct

# POLYTONE (.pol, prochazkaml) -- the PC-speaker sibling of MONOTONE played by
# the same PTPlayer. Support was dropped from the monotone plugin: the only
# public sample (PTPlayer's jakim.pol) is format version 2, while both our
# vendored engine and current upstream PTPlayer only parse version 1
# ("\x08POLYTONE\x01"). With no playable v1 fixture, .pol is documented here
# rather than advertised. Re-add to monotoneplugin supported_ext if v2 lands.
.pol

# game boy tracker / paragon 5 (s.hockenhull). .mgb = the editor's serialized
# project file ("GameBoy Music Module" tagged tree), NOT a gb rom -- so GME's
# game boy emu can't run it directly. Only the closed pt.exe reads .mgb; the
# open replayer (gbdev/GBSoundSystem) consumes the tracker's EXPORTED .z80/.inc
# driver data, not raw .mgb. No portable open parser anywhere (UADE/OpenMPT/
# libxmp/ZXTune/rePlayer/GME/vgmstream). RE-prototyped: container + note stream
# are decodable (plain binary, not compressed) but full support = leaf decode ->
# driver data -> RGBDS harness rom -> GME, with no headless reference to validate
# against the closed gui. Parked. Re-open if a portable .mgb player lands.
.mgb

# all sound tracker (.ast, modland "All Sound Tracker"). NOTE: .ast is shared and
# NOT fully unsupported -- UADE's "ActionAmics" eagleplayer plays the V0.1 binary
# replay dumps (pointer table + "ACTIONAMICS SOUND TOOL V0.1", e.g. dynablaster.ast),
# so no bare ".ast" line here. But the modland corpus is the tracker's NATIVE
# versioned save format (Pascal-string magic \x08"AST 00xx": "AST 0001","AST 0032"
# ...), a different layout the V0.1 testversion player cannot parse -- it loads,
# runs and emits silence while UADE reports success. No open replayer exists for
# the native format. UADEPlugin::canHandle declines it (isAllSoundTrackerNative)
# so those files Skip cleanly instead of false-positiving as a silent "ok". Re-open
# if a player for the "AST 00xx" native format lands.

# edlib packed (.edl, modland "Ad Lib/EdLib Packed", ~179 tunes; JCH/Vibrants'
# 1994 DOS AdLib/OPL2 tracker). EdLib has TWO formats: .d00 is the EXPORTED,
# documented, playable file (magic 'JCH'\x26\x02\x66, played by AdPlug's d00.cpp
# "Packed EdLib"), while .edl is the tracker's NATIVE editor PROJECT file --
# undocumented (the modland format doc covers only .d00) and not a disguised D00:
# flash2.edl has no 'JCH' magic, fails Cd00Player v0/v1 validation, and its
# verbatim-repeating 15-byte byte patterns rule out it being an aPLib/LZ-packed
# D00 (so adding .edl to adplugin supported_ext would NOT help). No open replayer
# or .edl decoder exists anywhere. The replayer for the OUTPUT already ships
# (AdPlug d00.cpp); the only gap is .edl->.d00 conversion, which EdLib 1.05a
# itself does deterministically. Fix = OFFLINE batch-convert the corpus to .d00
# (DOSBox EdLib, or RE the native format against the editor's in/out) then onboard
# the .d00 -- data-only, zero new in-app code. Parked; not worth an in-app decoder.
.edl

# forgotten worlds game music (.fw, magic "FWMP", modland "Forgotten Worlds/
# Mark Cooksey", ARC Developments / adapted by Psygore). NOTE: no bare ".fw"
# line -- UADE has a ForgottenWorlds_Game eagleplayer (prefix fw) that LOADS and
# runs, but it emits pure silence on all five modland tunes. Verified against
# upstream uade123 3.05: every .fw renders as silence (peak 0/0) while a TFMX
# module renders loud, so this is a player bug in UADE itself, not our wiring,
# and is unfixable without RE-patching the player binary. UADEPlugin::canHandle
# declines the FWMP magic (isForgottenWorlds) so they Skip cleanly instead of
# reporting NO SOUND. Re-open if a working ForgottenWorlds player lands upstream.
.fw

# picatune (.smufi, modland "Picatune", 20 tunes by BeRo/FreQvibez/Lyzzard).
# BeRo's (Benjamin Rosseaux) Picatune / Picatune 2 is a CLOSED-source freeware
# Win32 synth-sequencer; the .smufi ("Picatune Soundtrack") is an XML PROJECT
# file describing a softsynth patch graph -- <operator> nodes (master/delay/
# compressor/generator), <oscillator> waveforms with ADSR, and <order>/<pattern>
# sequencing. There is NO sample/PCM data: every voice is produced by Picatune's
# built-in DSP, so playback needs the closed engine. No open replayer exists
# anywhere (UADE/OpenMPT/libxmp/ZXTune/AdPlug/rePlayer none parse this XML synth)
# and the synth semantics (oscillator modes, waveform tables, x/y operator
# routing, phasefactor/color params) are undocumented -- same dead-end as
# .axs/.psy. Without this line each .smufi just downloads (FTP CODE 226) then
# stalls with no decoder. NOTE: unrelated to Shiru's ZX-beeper "Picatune2"
# (.pt2). Re-open if BeRo opens the engine or a portable player lands.
.smufi

# berotracker (.brt, modland "BeRoTracker", 19 tunes). BeRo's (Benjamin
# Rosseaux) own module tracker -- closed-source Delphi/FreePascal Win32 app,
# source never published (not among his ~50 github.com/BeRo1985 repos). The
# .brt format itself diverged from IT into its own engine and is undocumented;
# OpenMPT's own open feature request for it (bugs.openmpt.org #1483) has sat
# unresolved since the maintainer judged BRT "more advanced than OpenMPT in
# several ways," making any reimplementation "almost guaranteed to be a lossy
# conversion" -- their suggested fix was BeRo adding an in-app BRT->IT
# exporter, not something we control. No open replayer parses .brt anywhere
# (UADE/OpenMPT/libxmp/ZXTune/AdPlug/rePlayer). Same dead-end shape as BeRo's
# other closed format above (.smufi). Re-open if BeRo publishes the format or
# a portable player lands.
.brt

# atari exexutable, pruned from db but just in case:
.xex

# 0CC-FamiTracker (.0cc, BotB corpus). Carries the SAME "FamiTracker Module" magic
# and file-version 0x0440 as a plain .ftm, so the shipped FamiTracker CX engine's
# gate accepts it -- but 0CC's headline feature is multichip expansion the engine
# never implemented. Tested 20 real BotB .0cc files through famitracker-cx: only
# ~40% play; ~50% throw UNIMPLEMENTED_CX ("Instrument not supported" -- N163/
# INST_N106 and Sunsoft-5B/INST_S5B are commented out in FtmDocument.cpp), and
# ~10% HARD-SIGSEGV during render (CTrackerChannel::SetNote deref of a NULL channel
# at TrackerChannel.cpp:47 -- the module addresses N163 x8 / 5B channels the
# ChannelMap never allocated). The crash is a raw signal, NOT a catchable
# FtmDocumentException, so a bad .0cc takes down the whole app, and it fires mid-
# song (past load) so "loads OK" is no guarantee. Adding "0cc" to
# famitrackerplugin supported_ext was tested and REJECTED -- 40% playback isn't
# worth a >=10% process-crash rate. Proper support needs a 0CC-aware engine
# (N163/5B channels + expansion ChannelMap + extended blocks), a real upgrade.
# See botb-onboarding memory. Re-open if the engine is upgraded.
.0cc

# 0CC-FamiTracker compact export (.kftm, BotB corpus). NOT the "FamiTracker Module"
# block container at all -- magic is "kftm\x01\x00" followed by a 32-byte inline
# song name. base famitracker-cx has no reader for this format, and it shares the
# N163/5B expansion dead-end above. Would need its own parser AND a 0CC-aware
# engine. Parked with .0cc.
.kftm

# --- demoscene executables / disk images, 2nd pass (2026-07-14) ---
# Same class as the .ps1/.tic entries in the 2026-07-10 block above, found by
# re-auditing the whole live index (776,777 rows) for extensions that route to
# no plugin and are not listed here. All reached us through demozoo/scene.org
# compo-directory URL onboarding: they are ROMs, disk/tape images and native
# executables whose music is produced by running the machine code, not by
# decoding a module. Playing any of them means hosting the whole target system
# (a C64/CPU+VIC+SID, a 6502 NES, a 68k Genesis, an x86 PC), which no plugin in
# the stack does -- GME plays the *ripped* chip logs (nsf/gbs/vgm/sid), never
# the parent ROM or image. Each just downloads and then dead-ends in the GUI.

# c64 disk image (compo entries submitted as a whole .d64 disk, not a .sid rip)
.d64

# native executable music (DOS/Windows PE "executable music" compo entries)
.exe

# ines nes rom -- executable music as a cartridge image, not an .nsf rip
.nes

# sega genesis/mega drive rom -- likewise not a .vgm/.gym rip
.gen

# zx spectrum / c64 tape image (beeper + ay compo entries shipped as .tap)
.tap

# .tsd -- 26 files, two unrelated dead ends that share the extension: 24 are
# Fujiology Atari ST compo files (MUSIC/FT/TUNEZ/T/TAO/TSD/), and 2 are modland
# "TSS/- unknown/galg main.tsd|galg title.tsd" -- i.e. members of the TSS format
# dir, whose own .tss tunes are equally undecodable (no open TSS replayer in
# UADE/OpenMPT/libxmp/ZXTune/rePlayer). NOT a demoscene executable like the rest
# of this block; documented here because both variants are unplayable today.
# Re-open the modland pair if a TSS replayer ever lands (it would cover .tss too).
.tsd

# .spi -- 3 Fujiology files, all under MUSIC/FT/TUNEZ/G/GEM_TWO/EPSS/. Unidentified
# Atari ST "EPSS" format; no reader anywhere in the stack and no public spec found.
.spi

# --- MusicMaker V8 instrument packs indexed as songs (2026-07-14) ---
# modland "MusicMaker V8 Old/- unknown/" ships 7 files: 2 real tunes
# (best of guitars.sdata, moveback.sdata -- "sdata" is a UADE supported_ext,
# eagleplayer.conf "MusicMaker-8V prefixes=mm8,sdata") and 5 members of their
# 3-part instrument pack. UADEPlugin::getSecondaryFiles already pulls the pack
# in for the .sdata ("<name>.ip" samples + "<name>.ip.l"/"<name>.ip.n" metadata,
# see UADEPlugin.cpp ~line 826), so the packs must NOT be indexed as standalone
# songs -- they are not tunes and render nothing on their own. Skipping them here
# does not affect the .sdata tunes, which keep fetching all three parts.
# test.cpp's shouldIgnoreFile already treats ip/.ip. as aux for the same reason.
.ip

# The pack's other two members (".ip.l"/".ip.n") are deliberately NOT listed.
# They route on the last dot, i.e. "l" and "n", and single letters collide with
# Amiga prefix-form names whose TITLE happens to end that way: adding ".l" was
# tried and immediately dropped a real playable tune --
# "/Demo/Larsen_Sten/Egregious.lha/unic.take-no-fab by s.l.l", which routes on
# its "unic" prefix (UnIC Tracker, a UADE supported_ext) and only looks like a
# ".l" file to an extension matcher. ".n" measured clean today, but it is the
# same shape and one new title ending in ".n" would silently kill a good song.
# The 3 leftover .ip.l/.ip.n rows stay indexed; that cosmetic wart is a better
# trade than a routing regression. A companion-aware indexer rule (cf. test.cpp
# shouldIgnoreFile's ".ip." pattern) is the right fix if it is ever worth doing.

# liquid tracker (.liq, modland "Liquid Tracker/", 13 tunes; Nir Oren's mid-90s
# DOS tracker, magic "Liquid Module:"). TESTED AND REJECTED -- do NOT wire up
# libxmp's liq_loader. The loader exists and is already vendored under
# zxtune/3rdparty/xmp, and a slice plugin (the musxplugin/fnkplugin pattern)
# builds and plays: 10 of 13 tunes render audio. But its parse DESYNCS on 5 of
# the 13, tripping libxmp's own asserts:
#   liq_load.c:489  assert(b[] == "LDSS")            background/name/part of the sky
#   liq_load.c:209  assert(event->note <= 107 ...)   free mind
#   liq_load.c:443  assert(event->fxt <= 26)         free space
# Those asserts are LIVE in our build -- the tree overrides CMAKE_C_FLAGS and
# never defines NDEBUG (check: build.ninja FLAGS for any xmp slice has no
# -DNDEBUG despite CONFIG=Release) -- so each of those 5 files abort()s the whole
# app, not just the tune. Compiling the slice with -DNDEBUG "fixes" the abort but
# is worse: assert(event->note <= 107) and assert(event->fxt <= 26) guard ARRAY
# INDICES, so silencing them turns a clean abort into out-of-bounds reads with a
# desynced parse (and background.liq renders pure silence either way). Same call
# as .0cc above: partial playback is not worth a ~38% process-crash rate. Real
# support = fixing libxmp's liq loader upstream (the 1.xx instrument/pattern
# layout it mis-seeks), not wiring it in as-is. Re-open if libxmp fixes it.
# NOTE: Funktracker (.fnk) from the same triage was the opposite result -- 7/7
# tunes clean with asserts live -- and SHIPPED as fnkplugin.
.liq

# --- companion files indexed as standalone songs (2026-07-14) ---
# These are support files, never tunes: they carry no sequence data and render
# nothing on their own, but modland/AMP ship them next to the songs and the
# indexer listed them as rows. The plugins already fetch each one as a secondary
# file of the tune that needs it, so dropping the standalone rows costs nothing.
#
# SAFE ONLY BECAUSE THE GATE IS MULTI-AWARE: all three of these also appear as
# the FIRST member of MULTI: groups whose later members are real, playable tunes.
# songIsUnsupported() (MusicDatabase.cpp) skips a group only when EVERY member is
# unsupported, so the groups below survive while the standalone rows are dropped.
# Listing any of these under a first-member-decides gate would have deleted 451
# playable songs (74 Quartet .4v + 374 .minipsf2 + 3 .smus). Keep that in mind
# before adding any further companion extension here.

# quartet st sample bank ("SMP.set" / "<tune>.set"), paired with the .4v songs
# that quartetplugin plays. 149 standalone rows; the 9 "MULTI:.../SMP.set + .4v"
# groups keep working.
.set

# playstation 2 sound format shared driver/sample lib, referenced by the
# .psf2/.minipsf2 tunes in the same dir (aoplugin pulls it in as a psflib).
# Never a standalone row today -- all 6 rows are MULTI groups that survive -- so
# this line is pre-emptive: it stops a bare "*.psf2lib" from ever becoming an
# entry if a future rip lists one.
.psf2lib

# edlib d01 (.d01 song + .s01 instrument bank, modland "Ad Lib/EdLib D01/JCH/",
# 3 tunes = 6 rows). TESTED, DO NOT "just add d01 to adplugin supported_ext" --
# that would make canHandle claim the file and fromFile then fail, turning a
# clean Skip into a hard FAIL. AdPlug's Cd00Player rejects these two ways over,
# verified by linking libadplugin.a directly and calling Cd00Player::load():
#   crooner.d00         -> LOADED (control)    crooner.d01 -> REJECTED
#   crooner.d01 renamed to .d00 -> STILL REJECTED
# Why: D01 is NOT a renamed D00. Decoded against the real d00header (id[6] then
# type,version,speed,subsongs,soundcard) the marker byte differs --
#   .d00 = type 0x00 / version 2      (single file, instruments inline)
#   .d01 = type 0x80 / version 1      (song only)
#   .s01 = type 0x01 / version 0      (the external instrument bank)
# so the v2-4 branch (needs type==0) declines, and the fallback v0/v1 branch
# hard-requires a ".d00" extension *inside the library* (d00.cpp ~line 92); after
# a rename it passes that but then reads a d00header1 at offset 0, where "version"
# lands on the 'J' of the JCH magic (0x4A > 1) and rejects again. Not the
# "reheadered old-style song" case either -- there is no valid old header at 0x6b
# (reads version=219, i.e. song data). Cd00Player only plays single-file modules
# with instruments inline; D01 splits the ~20KB bank into the .s01, so real
# support means REing the D01 layout + external-bank resolution.
# ROI is poor: of the 3 tunes, "crooner" and "the alibi" ALREADY play from
# "Ad Lib/EdLib D00/JCH/" (both verified LOADED), so only "gorgeous" is net-new.
# Same dead-end shape as .edl above. Re-open if an open D01 reader ever lands.
.d01
.s01

# stale editor backups: modland's "IFF-SMUS/Dr. Awesome/*/[Awesome-N].SMUS.bak"
# and one chipmusic.org "dj-brama-style_kamu.mp3.bak" (an S3 backup artifact
# whose ".bak" tail hides the .mp3 from FFMPEGPlugin's ext gate, so it dead-ends
# instead of playing). The 3 SMUS groups keep their playable .smus members.
.bak

# --- demozoo archive contents, from the member peek (2026-07-15) ---
# The 17k demozoo "Demoscene" archive rows now carry the REAL inner format in
# their `ext` column (filter_demozoo_archives.py --classify reads each archive's
# member list over HTTP Range), which is what finally makes this list reachable
# for them: songIsUnsupported() matches `ext`, and every archive row used to say
# "zip". These are the formats left in archives holding nothing playable -- each
# verified to be claimed by NO plugin (cmtest priority_map), so nothing is being
# hidden that we could decode. The rows stay in demozoo.txt on purpose: adding a
# decoder later means deleting a line here + a reindex, not re-onboarding.
.adf
.bin
.com
.dll
.emd
.font
.iff
.j98
.mmd
.oss
.pp
.rbs
.sample
.t64
.wmv

# Archives whose only content is a readme -- ~100 scene.org compo entries are a
# lone .txt (occasionally .nfo/.diz): the tune was never uploaded, so the row is
# a dead search hit that errors with "No playable tracks in archive". The peek
# reports the readme as the row's `ext` (filter_demozoo_archives.py
# first_member_ext) precisely so these lines can gate them. Safe: no plugin
# claims txt/nfo/diz, and no collection uses them as an ext -- these lines can
# only ever match an archive that contains nothing else. Not deleted, so a
# re-upload upstream just needs a re-peek.
.txt
.nfo
.diz