@@ -0,0 +1,294 @@
+# RFC: nameparser 2.0 — a new core API
+
+Status: **open for feedback** — no deadline. Implementation may proceed
+in parallel; this document will be amended (as commits to this PR) as
+implementation teaches us things. Comment on specific lines here, or on
+the discussion issue (#284) for general reactions.
+
+## The short version, and the promise
+
+nameparser 2.0 introduces a new, immutable core API and keeps the
+existing one working:
+
+```python
+from nameparser import parse
+
+name = parse("Dr. Juan Q. Xavier de la Vega III (Doc Vega)")
+name.given        # "Juan"
+name.family       # "de la Vega"
+name.family_base  # "Vega"
+name.title        # "Dr."
+```
+
+**The compatibility promise:** code that runs warning-free on nameparser
+1.4 keeps running on every 2.x release, with identical results, getting
+at most `DeprecationWarning`s. `HumanName` continues to exist as a
+compatibility layer over the new core. The exceptions to "identical
+results" are (a) parse-output changes that are deliberate bug fixes,
+each documented in the release log with its issue, and (b) three narrow
+edge cases listed in the Migration section. Removal of the old API
+happens in 3.0, not 2.x, and only after warnings have named every
+replacement.
+
+If you only use `HumanName` and never customize configuration, 2.0
+should be a non-event for you. The rest of this document is for
+everyone else — and for anyone who wants to push on the new design
+while it is still cheap to change.
+
+## Why a new core
+
+Fifteen years of issues point at the same architectural limits:
+
+- **`HumanName` is four things in one mutable class** — parser, state,
+  vocabulary predicates, and formatter. Every feature grows all four at
+  once (it is currently ~1,600 lines), and mutation semantics are
+  inconsistent: assigning `full_name` re-parses, assigning `last` does
+  not, `capitalize()` mutates in place.
+- **Parse state is `list[str]`**, so position information is lost and
+  re-derived by value lookups. That is the structural root cause of a
+  whole family of prefix-join bugs (#100 and relatives): when the same
+  string appears twice, the lookup answers the wrong question.
+- **Configuration is a shared mutable module global** (`CONSTANTS`).
+  Most of `nameparser.config`'s ~850 lines exist to make shared
+  mutation survivable (custom set managers, descriptors, cache
+  invalidation), and it still causes test pollution and cross-thread
+  surprises.
+- **Equality is broken by design** (#223): case-insensitive equality,
+  equality with plain strings, and hashability are mutually
+  inconsistent promises. 1.3.0 deprecated `==`/`hash()` and shipped
+  `matches()`/`comparison_key()`.
+- **The roadmap needs structure v1 can't cheaply provide**: family-first
+  name order (#270), per-dataset configuration without global mutation,
+  and honest reporting of genuinely ambiguous parses.
+
+The rules-based approach is not changing. nameparser's niche is
+deterministic, auditable, zero-dependency parsing — the same input
+always parses the same way. The rewrite reorganizes *how the rules run*,
+not what the library is.
+
+## The new API
+
+### Parsing produces an immutable value
+
+`parse()` returns a `ParsedName`: a frozen dataclass. There is no
+mutation anywhere in the new API — editing is `replace()`, which
+returns a new value:
+
+```python
+name = parse("John Smith")
+name2 = name.replace(given="Jane")   # new object; `name` unchanged
+```
+
+`parse()` is **total over `str`**: any string — empty, emoji,
+garbage — returns a `ParsedName` rather than raising. Content problems
+are represented in the result; the only exceptions are type errors
+(`bytes` raises `TypeError` — decode first).
+
+Equality is strict structural equality, and the semantic comparisons
+are explicit and separate — `comparison_key()` for dedup/sorting and
+`matches()` for component-wise case-insensitive comparison — resolving
+the #223 trilemma by not conflating the three meanings of "equal".
+
+### Two access layers: strings for the 95% case, tokens for the rest
+
+The string properties behave like v1's: always `str`, `""` when empty.
+Underneath, `ParsedName` carries the actual parse: a tuple of `Token`s,
+each with its **span** (position in the original string), its role, and
+tags recording *how it was classified*:
+
+```python
+name = parse("Juan de la Vega")
+name.tokens_for(Role.FAMILY)
+# (Token('de' @5:7 FAMILY {particle}),
+#  Token('la' @8:10 FAMILY {particle}),
+#  Token('Vega' @11:15 FAMILY),)
+```
+
+This is the structural fix for the value-lookup bug family: pipeline
+stages address tokens by index, so "which 'de' did you mean?" cannot
+arise. It is also the debuggability story — "why is 'Van' in my family
+name?" is answered by the token's tags, and `repr(name)` prints a
+compact component-per-line breakdown like v1's.
+
+Derived views (`family_base`, `family_particles`, `surnames`,
+`given_names`) are pure filters over tokens — they can never disagree
+with the parse.
+
+### `given` and `family`, not `first` and `last`
+
+The v1 field names describe Western *position* — they date from before
+the parser even supported comma formats, when position was all there
+was. They read wrongly the moment name order varies: under family-first
+order, `name.last` would return the first word. 2.0 adopts the names
+the standards world already uses (HTML autocomplete
+`given-name`/`family-name`, vCard, Unicode CLDR):
+
+- `first` → `given`, `last` → `family`
+- `last_base`/`last_prefixes` → `family_base`/`family_particles`
+- config: `prefixes` → `particles` (every web standard uses "prefix"
+  for honorifics like "Mr." — v1's usage collides; the linguistics term
+  for van/de la is *particle*)
+- `title` and `suffix` **stay** — they have ecosystem gravity, and the
+  positional taxonomy ("pre-nominal"/"post-nominal") is documented
+  rather than spelled into identifiers.
+
+`HumanName` keeps every v1 spelling, including `string_format` keys, so
+existing code is unaffected; only new-API adopters see the new names.
+
+### Configuration: `Lexicon` (vocabulary) and `Policy` (behavior)
+
+v1's `Constants` mixes three kinds of thing; 2.0 splits them by what
+varies together:
+
+- **`Lexicon`** — vocabulary data (titles, suffixes, particles,
+  conjunctions, …). Immutable; composable by union; varies by
+  *language*.
+- **`Policy`** — behavior switches (`name_order`, patronymic rules,
+  delimiter pairs, …). Immutable; varies by *data source*.
+- **Rendering parameters** — `string_format`, initials style,
+  capitalization — move to render-time arguments; they vary by *output
+  destination* and are no longer configuration at all.
+
+```python
+from nameparser import Parser, Lexicon, Policy, FAMILY_FIRST
+
+parser = Parser(
+    lexicon=Lexicon.default().add(titles={"dra"}),
+    policy=Policy(name_order=FAMILY_FIRST),
+)
+parser.parse("Wang Xiuying").family   # "Wang"
+```
+
+A `Parser` is immutable, thread-safe, and does all its compilation once
+at construction — build it at startup, share it everywhere. There is no
+shared mutable configuration in the new API, which retires the entire
+test-pollution/config-war problem class. `name_order` is an order-spec
+tuple (`FAMILY_FIRST = (Role.FAMILY, Role.GIVEN, Role.MIDDLE)`) rather
+than a boolean, because Vietnamese order — family, middle, *then*
+given — needs a third arrangement (#270).
+
+### Honesty about ambiguity
+
+Some parses are genuinely undecidable by rules — "Kelly Michael" could
+be either order; in "Van Johnson", "Van" is a Dutch particle *and* a
+common given name. v1 silently picks a reading. 2.0 picks the same
+reading **and says so**:
+
+```python
+parse("Van Johnson").ambiguities
+# (Ambiguity(PARTICLE_OR_GIVEN, "leading 'van' can be a particle or a
+#   given name; chose given", tokens=(...)),)
+```
+
+`Ambiguity.kind` values are stable API you can filter on. This is also
+groundwork: a future `parse_all()` returning ranked alternative
+readings, and `explain()` returning a stage-by-stage parse trace, are
+designed for (the pipeline supports both) but deliberately not in 2.0.
+
+### Locales: opt-in packs, never auto-detection
+
+Language-specific behavior cannot be auto-detected — "Tham Jun Hoe" is
+token-for-token indistinguishable from a Western three-token name, and
+"Ali" is Arabic or Italian. So locale support is explicit:
+
+```python
+from nameparser import parser_for, locales
+
+parser_for(locales.RU).parse("Ivanova Anna Sergeevna").given   # "Anna"
+```
+
+A locale pack is pure data: a vocabulary fragment plus a partial policy
+patch, folded in at parser construction. 2.0.0 ships `RU` and `TR_AZ`
+(formalizing the patronymic support added in 1.3.0); Chinese/Korean
+(bundled surname lists + segmentation of unspaced input), Vietnamese,
+and Japanese (via a `nameparser[ja]` extra with a pluggable segmenter)
+are designed and staged for 2.x minors. Vocabulary that cannot misfire
+on other languages' names (e.g. Cyrillic honorifics) goes in the
+default lexicon instead of packs.
+
+## Migration: what actually happens to existing code
+
+**`HumanName` in 2.0** is a wrapper over the new core with v1's exact
+mutation semantics, field spellings, and formatting attributes. The v1
+test suite runs against it as the compatibility gate. v1 pickles load
+throughout 2.x.
+
+**`CONSTANTS` keeps working** — mutations are honored exactly (the
+facade rebuilds its config snapshot when the shared config changes) and
+emit a `DeprecationWarning` pointing at the replacements. Mutating a
+*private* `Constants` instance does not warn.
+
+**Already-scheduled removals land as warned** (every one was deprecated
+in 1.3.0 or 1.4 first): `==`/`hash()`, bytes input, slice item access,
+`empty_attribute_default`, `constants=None`, silent config-typo
+fallbacks, legacy pickle blobs, `add_with_encoding()`. Python floor
+becomes 3.11; the last dependency is dropped.
+
+**The three narrow behavior exceptions** (everything else identical):
+
+1. `*_list` attributes return snapshots — in-place
+   `name.first_list.append(...)` no longer mutates parse state
+   (assignment still does).
+2. Subclasses overriding parsing hooks (`parse_pieces`, `is_title`, …)
+   are detected and warned at construction: the facade delegates to the
+   new core, so such overrides are no longer called (#280). If you do
+   this, we want to hear what your override does — that is a feature
+   request we would rather absorb than break.
+3. Assigning to `CONSTANTS.regexes` raises with a pointer to the named
+   `Policy` flag that replaces it. Custom regex injection has no
+   equivalent in the new model; nothing is silently ignored.
+
+**Timeline:** 2.0.0's only *new* warning is on shared-`CONSTANTS`
+mutation. A warning on `HumanName` construction itself comes in a later
+2.x minor, once the new API has proven out — and 3.0 (which removes the
+facade) does not ship before that warning has been out in a release.
+
+## Alternatives considered and rejected
+
+- **Auto-detecting name language/order** — impossible in principle, not
+  just hard; see the Locales section.
+- **A statistical/ML core** — determinism and auditability are the
+  point of this library; the token/ambiguity model leaves room for
+  optional escalation later without changing the core.
+- **Keeping `first`/`last`** — see the naming section; the facade keeps
+  them for v1 code, which serves the familiarity constituency without
+  wiring position-based names into the new API.
+- **A boolean `family_name_first`** — cannot express Vietnamese; the
+  order-spec tuple can (#270).
+- **Renaming `title`/`suffix` to `pre_nominal`/`post_nominal`** — more
+  precise, but no ecosystem uses those as field names; precision lives
+  in the docs instead.
+- **Vocabulary entries as records with per-word metadata** — rejected
+  under "no vocabulary data without a consuming rule": a flag the
+  parser doesn't read is a promise it doesn't keep. Cross-category
+  words are handled by dual set membership (closed classes) or the two
+  `_ambiguous` sets (closed-vs-open), each of which has a consuming
+  rule.
+- **A user-replaceable regex escape hatch** — replaced by named policy
+  flags; anything not expressible through them is a feature request we
+  want to see.
+
+## Feedback we are specifically looking for
+
+No deadline — comment whenever, on lines here or on the discussion
+issue. The questions where answers would most change things:
+
+1. **Do you compare `HumanName` objects with `==` (especially against
+   plain strings), or use them in sets/dicts?** 2.0 changes this to
+   object identity (warned since 1.3.0; `matches()`/`comparison_key()`
+   are the replacements). This is the one silent behavior change in the
+   release.
+2. **Do you subclass `HumanName` and override parsing methods?** Tell
+   us what your override accomplishes — we would rather absorb it as a
+   feature than break it silently (#280).
+3. **Do you mutate `CONSTANTS` after startup** (not just at import
+   time)? Anything beyond startup-time customization we should know
+   about?
+4. **Do you assign custom regexes to `CONSTANTS.regexes`?** What for?
+   Each real use case either becomes a named policy flag or stays
+   broken — we would like the list to be complete before 2.0.
+5. **`given`/`family` instead of `first`/`last`** in the new API: does
+   this help or annoy you?
+6. **Is there anything you build on top of nameparser** (wrappers,
+   pipelines) where the compatibility promise as stated would still
+   break you?