Hopefully not too late to the party, but here are my votes. Like Richard’s, anything other than “yes” or “no” is the reason why it’s a “no”.

Module System
=============

ImportQualifiedPost: yes
PackageImports: yes
NoImplicitPrelude: a very breaking change

Notation
========

BlockArguments: yes
MultiWayIf: yes
LambdaCase: yes
BinaryLiterals: yes
HexFloatLiterals: yes
NumericUnderscores: yes
NumDecimals: yes
OverloadedStrings: more problems along the lines of `show 1` that seem to trip up a lot of newcomers.
OverloadedLists: same as OverloadedStrings
OverloadedLabels: not a no as much as a "no strong opinion" - personally never preferred it over alternatives.
EmptyCase: yes
PostfixOperators: yes
LexicalNegation: yes
UnicodeSyntax: yes
NegativeLiterals: yes / N/A
TupleSections: yes
ImplicitParams: not obvious syntax, turning it on in cabal files/pragmas gives unfamiliar readers something to search for
ParallelListComp: yes
RecursiveDo: no strong opinion, I’ve never needed/used it in practice
TransformListComp: adds more complexity to list comp notation, which is already complexity on do-notation
Arrows: no strong opinion, but I’d lean towards no for the same reason as ImplicitParams
ApplicativeDo: probably quite a breaking change in practice
QualifiedDo: no
MonadComprehensions: no
NondecreasingIndentation: I had no idea about this until I saw the user guide example - this is probably unintuitive
RebindableSyntax: definitely a breaking change, not least because it implies NoImplicitPrelude
ExplicitNamespaces: yes

Data Types
==========

DatatypeContexts: no
ExistentialQuantification: yes
EmptyDataDecls: yes
RoleAnnotations: yes
StrictData: no, definitely a breaking change
GADTSyntax: yes
GADTs: the implied MonoLocalBinds probably means this would be a breaking change

Patterns and Guards
===================

BangPatterns: yes
ViewPatterns: yes
PatternSynonyms: no
NoPatternGuards: no
NPlusKPatterns: no

Records
=======

NamedFieldPuns: yes
RecordWildCards: yes, but a weak yes
DisambiguateRecordFields: yes
DuplicateRecordFields: yes
NoTraditionalRecordSyntax: no

Deriving
=======

DeriveGeneric: yes
DeriveLift: yes
DeriveDataTypeable: yes
EmptyDataDeriving: yes
StandaloneDeriving: yes
DeriveFunctor: yes
DeriveFoldable: yes
DeriveTraversable: yes
DerivingStrategies: yes
DerivingVia: yes
GeneralisedNewtypeDeriving: yes
DeriveAnyClass: no, I can imagine this would be rather dangerous

Class System
============

MultiParamTypeClasses: yes
NullaryTypeClasses: yes (albeit implied)
ConstraintKinds: yes
TypeSynonymInstances: yes
FlexibleInstances: yes
FlexibleContexts: yes
ConstrainedClassMethods: yes
DefaultSignatures: yes
InstanceSigs: yes
ExtendedDefaultRules: leads to some unexpected error messages
FunctionalDependencies: yes
QuantifiedConstraints: probably too new
UndecidableInstances: noooo
IncoherentInstances: noooo
UndecidableSuperClasses: noooo
OverlappingInstances: noooo

Types
=====

RankNTypes: yes
StandaloneKindSignatures: yes
KindSignatures: yes
LiberalTypeSynonyms: yes
ScopedTypeVariables: yes
ExplicitForAll: yes
AllowAmbiguousTypes: the class-site error is more useful than the use-site
ImpredicativeTypes: not yet
MonoLocalBinds: definitely a breaking change
NoMonomorphismRestriction: I’d lean weakly towards “yes”, but happy to be overruled.
PartialTypeSignatures: no
NamedWildCards: yes
LinearTypes: no
TypeApplications: yes
PolyKinds: yes
TypeOperators: yes
StarIsType: no, or even No
TypeFamilies: yes
TypeFamilyDependencies: yes
DataKinds: yes

FFI
===

ForeignFunctionInterface: no
CApiFFI: no
GHCForeignImportPrim: no
InterruptibleFFI: no
UnliftedFFITypes: no
StaticPointers: no

Low Level
=========

UnboxedSums: yes
UnboxedTuples: yes
MagicHash: yes
UnliftedNewtypes: yes

Macros
======

CPP: no
TemplateHaskell: yes
TemplateHaskellQuotes: yes
QuasiQuotes: yes


Other
=====

Unsafe: no
Safe: no
Trustworthy: no
Strict: no

Obsolete/Deprecated
===================

CUSKs: no
TypeInType: no
MonadFailDesugaring: no



If anything here seems contradictory, I can try to justify my answer.
My general rule was, “if it’s a battle-tested extension to the language rather than a change, and it’s reasonably google-able/popular, it’s probably fine”.

Thanks,
Tom