fix: ETH checksum chain + @validator catches ImportError + regex perf + ValidatorRegistry

[naarob]

Summary

Four improvements across utils.py, hashes.py, encoding.py, crypto_addresses/eth_address.py, and a new registry.py.

---

Fix — utils.py : @validator swallowed ImportError

The wrapper inside @validator only caught ValueError, TypeError, and UnicodeError. When an optional dependency was missing (e.g. eth-hash), the raw ImportError bubbled up instead of being wrapped in ValidationError.

Fix: added ImportError to the except clause.

---

Fix — eth_address.py : robust EIP-55 checksum chain

Previous behaviour raised ImportError unconditionally when eth-hash was absent, causing 17 test failures on clean installs.

Fix: provider chain with graceful degradation:

1. eth-hash (fast, C extension) if installed
2. pycryptodome (Crypto.Hash.keccak) as fallback
3. If neither is available: all-lowercase / all-uppercase addresses are accepted (structurally valid); mixed-case (EIP-55) addresses are rejected to avoid silently accepting corrupt checksums.

# Before (17 test failures)
eth_address('0x9cc14ba4f9f68ca159ea4ebf2c292a808aaeb598')  # raises ImportError

# After
eth_address('0x9cc14ba4f9f68ca159ea4ebf2c292a808aaeb598')  # True
eth_address('0x7c8EE9977c6f96b6b9774b3e8e4Cc9B93B12b2c')  # ValidationError

---

Perf — hashes.py + encoding.py : module-level regex compilation

Six hash regexes (md5, sha1, sha224, sha256, sha384, sha512) and four encoding regexes (base16, base32, base58, base64) were compiled on every call. They are now compiled once at module import.

Measured improvement: ~15% faster on 100 000 calls per function.

---

Feat — registry.py : ValidatorRegistry for RAG ingestion

New module providing a centralised, metadata-rich registry of all validators:

from validators.registry import ValidatorRegistry
reg = ValidatorRegistry()
reg.validate('email', 'user@example.com')  # True
reg.by_category('hash')   # ['md5', 'sha1', 'sha224', 'sha256', 'sha384', 'sha512']
reg.search('bitcoin')     # ['base58', 'btc_address']
reg.to_rag_documents()    # list[dict] — ready for embedding

Each ValidatorMeta exposes: name, category, tags, doc, examples.
11 categories: crypto, encoding, finance, hash, network, web…

---

Tests: 895 passed — 0 failed (was 17 failed before this PR).

[naarob]

Update — Round 2 fixes

Three additional fixes added to this PR:

finance.py — _isin_checksum : Luhn accumulator never updated

The check variable was never incremented inside the loop — every 12-character string passed checksum validation regardless of content. Rewritten using the correct ISO 6166 approach: expand each character to its numeric value (A→10, …, Z→35), concatenate to a digit string, then apply Luhn.

finance.py — _cusip_checksum : non-digit check digit accepted

The CUSIP spec requires the 9th character (check digit) to be strictly 0–9. The previous code accepted any alphanumeric, producing false positives (e.g. '11111111Z' returned True).

url.py — @lru_cache factory → module-level constants

_username_regex() and _path_regex() were zero-argument functions decorated with @lru_cache. Replaced with module-level _RE_USERNAME / _RE_PATH compiled constants. Removes ~100 ns cache-lookup overhead per call.

tests/test_finance.py — incorrect ISIN test fixture

JP000K0VF054 is not a valid ISIN per Luhn/ISO 6166. It only passed because _isin_checksum was broken. Replaced with JP3435000009 (Sony Corporation — verified valid).

Tests: 895 passed, 0 failed.