Fix IPv6 authority parsing: Host()/Port() return garbage, scheme-less IPv6 breaks

[prorochestvo]

Summary

Parse mangles every IPv6 authority. The instance splitter in parseAuthority
uses strings.SplitN(instance, ":", 2), which splits on the first colon — and
an IPv6 literal is nothing but colons. As a result Host() and Port() return
garbage for bracketed IPv6 hosts, and scheme-less IPv6 is force-dumped into
Database().

A detailed, executable implementation plan lives in the repo at
plans/004-ipv6-authority-parsing.md.
This issue tracks the work; the Open Questions below must be settled before
implementation starts.

Current behavior (verified against HEAD)

Input	Host()	Port()	Database()	Addr()	Marshal
redis://[::1]:6379	[	:1]:6379		[::1]:6379	redis://[::1]:6379
redis://[::1]:6379/0	[	:1]:6379	0	[::1]:6379	redis://[::1]:6379/0
redis://[::1]	[	:1]		[::1]	redis://[::1]
postgres://u:p@[2001:db8::1]:5432/db	[2001	db8::1]:5432	db	[2001:db8::1]:5432	round-trips (string)
http://[fe80::1%25eth0]:8080	[fe80	:1%25eth0]:8080		[fe80::1%25eth0]:8080	round-trips
[::1]:6379 (scheme-less)			/[::1]:6379		////[::1]:6379
[2001:db8::1]:5432/db (scheme-less)			/[2001:db8::1]:5432/db		////[2001:db8::1]:5432/db
redis://127.0.0.1:6379 (IPv4 baseline)	127.0.0.1	6379		127.0.0.1:6379	ok

Two distinct failure classes:

1. Authority form (redis://[::1]:6379): Host()/Port() are garbage, but
   Addr()/Marshal come out correct by accident — the wrong split
   (host="[", port=":1]:6379") is exactly reversed by Addr()'s
   host + ":" + port rejoin (the same mechanism as the Telegram token trick). The
   broken observable surface is Host()/Port().
2. Scheme-less form ([::1]:6379): no accidental save — the classifier sees the
   first colon (inside the brackets), fails the colon-then-digit test, and routes
   the whole literal into Database().

Root cause

• parseAuthority: strings.SplitN(instance, ":", 2) is bracket-naive.
• Parse scheme-less classifier: the first-colon heuristic is bracket-naive.
• Addr(): host + ":" + port does not bracket an IPv6 host.

Proposed approach (surgical, zero-dependency)

Fix in place; do not adopt net/url for this (that is a separate, larger
decision — see OQ-5). Roughly ~15 lines plus a test table:

1. splitHostPort (new unexported helper): a leading [ marks an IPv6 literal —
   the host runs to the matching ], with an optional :port after it. Without a
   leading [ it keeps the exact current first-colon split, so the Telegram trick
   and all existing cases stay byte-for-byte unchanged.
2. Addr(): bracket the host when it contains a : and is not already
   bracketed. assembleDSN / Marshal / Redacted route through Addr(), so they
   need no change (one home for the bracketing decision).
3. Classifier: one new case for a leading [ → reshape to the // authority
   form and reuse splitHostPort.
4. Tests: full IPv6 truth table (compressed/full/mixed forms, with/without port,
   zone id, IPv4-mapped, /db, ?opts, credentials, scheme-less) + round-trip
   assertions + regression set (existing cases + Telegram trick + IPv4).

net.SplitHostPort / net.JoinHostPort were considered and rejected: the former
rejects a missing port and the Telegram first-colon trick; the latter mandates a
port and would emit [::1]: for the host-only form.

Backward-compatibility (hard constraints)

• All existing TestParse success cases pass unchanged.
• The Telegram Addr()-as-token trick stays intact (bracket logic fires only on a
  leading [).
• No new dependencies, no CGO, regex preserved.
• Host()/Port() moving from garbage to correct is a behavior change; no known
  consumer feeds an IPv6 DSN today, but classify the commit fix: and flag it.

Open questions (decide before implementing)

• OQ-1 (headline): store the host unbracketed (Host() → ::1, matches
  net/url.URL.Hostname(); requires the Addr() bracketing edit) vs bracketed
  (Host() → [::1]; smaller diff, but a leaky accessor a caller must strip before
  net.Dial). Lean: unbracketed — the accessor is the public contract.
• OQ-2: zone ids ([fe80::1%eth0] / %25) and IPv4-mapped ([::ffff:1.2.3.4]):
  preserve the literal, percent-decode, or validate? Lean: preserve literal, no
  decode, no validation.
• OQ-3: Addr() with no port → [::1] or ::1? Near-forced [::1], else
  Marshal produces the invalid redis://::1.
• OQ-4: also normalize the host on SetHost for a path-independent Host()?
  Lean: yes, but optional under the unbracketed baseline.
• OQ-5: keep this surgical now, or fold IPv6 into the net/url engine decision?
  Lean: fix IPv6 surgically now; let the net/url migration proceed on its own
  timeline.

Full task breakdown, acceptance criteria, and the complete 13-row truth table are
in plans/004-ipv6-authority-parsing.md.