[ilyam8]

Summary

Phase 6 (guardrails), codec reads. No behavior change, no exported API change.

• Header fields. Twenty header field reads each spelled seven lines (read, wrap the error as "error parsing
  ", assign when the field has the expected type). A small fieldReader keeps the first error, with the same
  text, and makes every later read a no-op, so each field is a three-line read-and-assign and a decoder checks the
  error once. Tried on unmarshalResponse first (which now takes the PDU type from its caller, and whose GetBulk/else
  branches hold two reads each), then used for the community, the SNMPv1 trap header, the SNMPv3 header, the scoped
  PDU and the USM parameters. readField is its one read; readInt, readString and readUint are gone, and the
  version and the varbind name, single reads each followed by its own check, call readField directly.
• Varbinds. The varbind list loop's body is readVarbind, the read-side counterpart of appendVarbind; no reader
  or tag variable is shadowed any more.
• Version. readVersion takes only the message. Its predecessor took a packet only to nil-check it (no caller
  passes nil) and to preset its variables, which unmarshalHeader now does itself under the same condition, marked as
  the known bug it is. unmarshalPayload loses the same unreachable nil check.
• Trap table. UnmarshalTrap's table branch is unmarshalTrapByUser: one attempt per parameter set on a clone of
  the trap, with an explicit last error instead of one carried by a variable shadowing the named result.

Validation

• New pins, each passing on the code before the refactor: every touched read failing in turn, with its exact error
  text and the partial packet SnmpDecodePacket returns (the decode golden records outcomes only); the trap table's
  attempts (a later set decodes after an earlier one failed, the last attempt's error is reported, the caller's trap
  is untouched); a table receiver decoding SNMPv1 and SNMPv2c traps; mistyped header fields keeping the value already
  there (the USM parameters start as a copy of the decoder's).
• Codec, engine, FIPS 140-only and USM goldens unchanged.
• 39 mutation probes on the moved and rewritten code: 36 killed (the new pins caught three: a table receiver decoding
  SNMPv1/SNMPv2c traps, and mistyped header fields keeping the value already there); 3 cannot change any reachable
  outcome.
• A/B (8 interleaved rounds against the previous PR's head): allocations unchanged (SNMPv3 trap decoding one fewer);
  time +1-3% on larger messages and up to +9% (+28 ns) on the smallest SNMPv3 one, about 2.3 ns per header field read:
  a reader that skips after a failure and keeps the error does not fit Go's inlining budget. Accepted for the
  readability. A first form of the reader kept the reader beside its error, and escape analysis, which tracks a struct
  as one value, moved every decoder's reader to the heap (+2 allocations per message); fixed here.
• go test -race, FIPS 140-only, golangci-lint, cross-builds, govulncheck: pass locally.