Bug-fix pass: ~30 verified fixes from public-API stress test (prep for 0.1.3)

CHANGELOG.md

@@ -1,5 +1,85 @@
 # Changelog
 
+## 0.1.3 (2026-06-11)
+
+Bug-fix release from a systematic stress-test of the public API surface. ~30
+verified fixes, no new features.
+
+**Resolution correctness.** Dotted abbreviations are no longer misclassified
+as missing-value markers: `"U.S.A."` resolves to `country/USA` (it previously
+resolved to an unrelated org entity) and `"U.K."` to `country/GBR`, while
+genuine null markers (`#N/A`, `--`, `.`) still return no match. Mixed-case
+inputs (`"fRaNcE"`, `"SUDan"`) resolve like their standard casings.
+Zero-padded ISO numeric codes (`"004"`) now resolve, `to="iso_numeric"` emits
+the canonical zero-padded form, and the pycountry-style `numeric` alias works
+(`entity("France").numeric` → `"250"`). `snap()` accepts free-text candidate
+labels as documented, alongside entity IDs. Punctuation-only inputs (`"."`,
+`"?"`) return no match instead of an internal error. `EntityRecord.aliases`
+no longer leaks the canonical name or duplicates.
+
+**Crashes.** `bulk()` no longer crashes on polars Series input without a
+pivot; `output="record"` builds primitive records that `to_polars()` accepts.
+`ResolutionResult` pickles cleanly.
+
+**Validation and errors.** Enum-like parameters are validated eagerly with
+did-you-mean suggestions instead of silently accepting typos: `on_ambiguous`
+(`resolve_id`), `on_missing`/`on_error`/`on_ambiguous` (`bulk`), `default_to`
+types (`configure`), `confidence_threshold` and `domain` (`parse`),
+`name:` language segments, and `ResolutionContext.country`.
+`AmbiguousResolutionError` hints are candidate-aware (no longer suggesting
+`entity_types=` when the tied candidates share one type) and `str()` previews
+the top candidates. `to=<typo>` errors suggest the closest code system
+instead of dumping all of them, and the `domain=`-with-auto-routing error no
+longer references internals. `from_records()` reports the offending row and
+column for empty name cells.
+
+**Behavior consistency.** `configure()` no longer clears settings that are
+omitted from the call; passing `None` explicitly resets a setting to its
+default (`cache_dir`, `default_to`). Mutating a returned result's lists no
+longer corrupts the query cache. `as_of=` accepts ISO date strings on
+`members_of`/`is_member`/`related`/`within`. `bulk()` pandas output preserves
+`None` under pandas 3. `available_entity_types()` returns the fine-grained
+types that `entity_types=` accepts. BYOD labels containing
+NFKC-compatibility characters (`™`, `№`) now round-trip; existing BYOD disk
+caches are rebuilt automatically.
+
+**API consistency.** Scalar `resolve()`/`resolve_id()` now coerce numeric
+input the same way `bulk()` does (`resolve_id(840)` → `country/USA`; integral
+floats like `840.0` from numeric dataframe columns coerce cleanly in both
+paths), and non-string types raise the `TypeError` the docstring always
+promised — `bool` included. `ResolutionContext(country=...)` accepts ISO
+alpha-3 alongside alpha-2 (`"USA"` and `"US"` behave identically). The pandas
+and polars accessors no longer convert caller mistakes into all-`None`
+columns: parameter-validation errors propagate (polars previously garbled
+them through `map_batches`), and `on_error` is exposed with the same
+`"raise"` default as `bulk()`. `ResolutionResult.reasons`, `.candidates`,
+and `.refinement_hints` are tuples now — the documented frozen contract is
+real, not advisory. Commonly raised errors (`UnknownCodeSystemError`,
+`UnknownOutputError`, `UnknownDomainError`, `OutputMissingError`,
+`DataPackNotAvailableError`, `CrosswalkError`, `ExplainNotAvailableError`,
+`NoModulesInstalledError`) are importable from top-level `resolvekit`;
+`resolvekit.errors` remains the canonical home.
+
+**Docs.** Corrected the `snap()` candidate guidance and the
+`UnknownOutputError`/`UnknownCodeSystemError` reference entries; refreshed
+stale confidence figures in the tutorials; documented that code
+auto-detection is case-sensitive by design while `from_system` is
+case-insensitive.
+
+**Performance.** The SymSpell typo index is now built in a background daemon
+thread during `Resolver` construction (default on), so the build cost no longer
+lands on the first query that passes the exact-match tiers. Opt out with
+`warm=False` on any constructor (`Resolver.auto(warm=False)`,
+`Resolver.from_modules(warm=False)`, etc.) to keep construction fully lazy.
+`resolvekit.warm()` and `Resolver.warm()` are new synchronous, idempotent,
+thread-safe functions that build all lazy indexes and return when they're ready
+— for servers or batch jobs that want deterministic readiness. The large-tier
+SymSpell index (706k terms, ~6 s to build on remote-data installs) is now
+cached as a locally-generated pickle under `<cache-dir>/compiled/` after its
+first build (~150 MB, loads in ~1.4 s on subsequent processes), keyed by the
+dictionary files and symspellpy version; existing bundled-only installs are
+unaffected.
+
 ## 0.1.2 (2026-06-11)
 
 **Fixed.** `download()` crashed on a clean install with "Missing package