Stop re-validating addresses the decoder just generated

[EONRaider]

Summary

On the decode path a MAC rendered from six raw bytes by bytes_to_mac() was immediately matched against mac_regex to prove what the conversion already guaranteed — and the same for validate_ipv4_addr() against inet_ntop() output. With #82 and #83 merged, re.Pattern.match was the largest single remaining cost in a corpus profile.

Ethernet, ARP and IPv4 now build their decoded instance directly (object.__new__ + object.__setattr__ per field), skipping __init__/__post_init__on that path only.

Strictness is unchanged — that was the thing to get right

The issue is explicit that construction-time strictness is a deliberate differentiator, so the invariant is pinned from both sides:

• Ethernet(dst="nonsense", ...) still raises InvalidMACAddressError; same for ARP and IPv4. Tested.
• Ethernet.decode(...) runs no regex at all — tested by patching the compiled patterns with a spy whose .match raises, so any re-validation fails the test loudly rather than silently costing time.
• Decoded instances are compared against constructed ones (equality touches every field), so a field the bypass forgot to set surfaces as an AttributeError in the suite instead of lurking.

The __post_init__ checks the bypass skips are ones decode() already establishes — IPv4's IHL is 4 bits and was rejected below 5, and the options are sliced to exactly ihl * 4 bytes after the buffer was confirmed to hold them. That reasoning is recorded at each site and in a new "Decode-path construction" section in _base.py.

Scope was kept deliberately narrow: the other protocols' __post_init__ are cheap integer comparisons, so bypassing them would buy little and risk more.

Measured

200k calls, best of run, CPython 3.12 on this machine:

	before	after
Ethernet.decode	2330 ns	830 ns	2.8×
IPv4.decode	4936 ns	2596 ns	1.9×
corpus walk (97 frames)	76,500 f/s	113,200 f/s	1.48×

That is ~3.8 µs saved per Ethernet+IPv4 frame, against the ~2.3 µs the issue attributes to the regexes alone — skipping the whole constructor path rather than only the validators accounts for the difference. Cumulative over #82 + #84: 58,300 → 113,200 f/s (1.94×).

Of the two options the issue weighed, this is the first (object.__new__); a central helper taking **fields measured 855 ns against 393 ns inline, because the per-field loop dominates, so the shortcut is written out at the three sites instead.

Verification

• uv run ruff check and uv run ruff format --check are clean
• uv run mypy is clean (strict)
• uv run pytest — 753 passed
• CHANGELOG.md entry under ## [Unreleased]

Closes#84.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QJnVMNGwTRDktC4rkABtgt

---

Generated by Claude Code