Cache DNS section parsing instead of re-serializing per access

[EONRaider]

Summary

layer7/dns.py called bytes(self) — a full re-serialization of the whole message — from six sites, because compression offsets are message-relative while the parser only held sections. One .answers access rebuilt the message 14 times. And answers / authorities / additionals each re-parsed all three sections, so reading all three parsed the message three times over.

Both of the changes the issue asks for, in order:

1. No more round-tripping. The parsers are module-level functions working directly from the sections bytes the instance already holds, in section-relative offsets. Only a compression pointer needs translating (subtract the 12-byte header length) — and a pointer addressing the header itself now lands below zero and raises InvalidFieldError instead of decoding header bytes as labels. bytes(self) appears in no accessor.
2. One parse for three sections, via a module-level lru_cache keyed by the immutable section bytes plus the counts.

The frozen-instance question

Nothing is stored on the instance: no object.__setattr__, no extra field, no cached_property (which slots=True rules out anyway). Hashability and equality are untouched, bytes(decode(x)) == x still holds, and the memo is bounded at 128 messages. Because it is keyed by value, two equal messages share one parse — records are frozen dataclasses, so sharing them between callers cannot leak mutation, which the tests assert directly.

Measured

Corpus DNS response, 20k iterations, CPython 3.12 on this machine:

	before	after
one uncached parse	17.03 µs	12.17 µs	1.4× (re-serialization alone)
.answers	16.84 µs	0.20 µs
.answers + .authorities + .additionals	52.08 µs	0.54 µs	ratio 2.9× → 1.03×

Corpus throughput is unchanged (~115k f/s) because the chain walk never touches record accessors — this is a win for callers that read records, which is what the accessors are for.

Verification

• Reading all three sections parses once — asserted via cache_info() (1 miss, 2 hits)
• bytes(self) is not called from any accessor — no occurrences remain in the module
• Round-trip and DNS corpus assertions unchanged
• uv run ruff check / ruff format --check / mypy clean; uv run pytest — 757 passed

Notes

One deliberate behaviour change worth a reviewer's eye: a compression pointer into the fixed header previously produced garbage labels and now raises InvalidFieldError. It is unreachable in well-formed messages, the fuzz contract (ProtocolError or a value, never anything else) is preserved, and there is a test for it.

Closes#85.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QJnVMNGwTRDktC4rkABtgt

---

Generated by Claude Code