Keep the hand-rolled Psych parser; pin the yaml.v3-incompatible invariants

[tannevaled]

Verdict: KEEP (control-run proven)

This branch set out to base the Ruby object model + byte-exact MRI-Psych emit on
the reference gopkg.in/yaml.v3 parser (Node tree), shedding the hand-rolled
parser — the same idea as go-ruby-hcl2. As instructed, it was gated control-first.
The control run says do not swap, and this PR ships the KEEP outcome instead of
a refactor.

What was measured (yaml.v3 v3.0.1)

A prototype fed the reference parser's Node tree into this package's own scalar
resolution and !ruby/* tag mapping (never yaml.v3's Go-typed Decode, which
collapses Float/Integer and uses YAML-1.2 core-schema resolution), then replayed:

1. The emitter corpus — every value Dump supports, parsed back through both
   loaders. Result: 100% parse-equivalent (30/30). yaml.v3 correctly reads
   everything this emitter writes.
2. The full yaml/yaml-test-suite (402 accept/reject tests). Result: 115
   verdict changes — 61 regressions against the required Psych verdict vs 54 fixes.
   Net conformance loss (335 → 328) that also breaks the shrink-only ratchet.

Why it cannot move (bidirectional, first-hand)

Invariant	This package	yaml.v3 v3.0.1
Float 2.0 vs Integer 2 emit	2.0 / 2	Marshal(2.0)==Marshal(2)=="2" (collapses)
:name scalar	Symbol("name")	String(":name") (no Symbol)
!ruby/object: / !ruby/range	*Object / *Range	not reconstructible
%YAML 1.2 directive (27NA)	accepts → "text"	rejects ("incompatible document")
nil-keyed explicit mapping (2JQS)	accepts	rejects ("did not find expected key")
stray flow close [ a, b, c ] ] (4H7K)	rejects	accepts
junk after flow (KS4U)	rejects	accepts

The 61 regressions split 50 accept-axis (Psych YAML-1.1 leniencies yaml.v3 rejects)

• 11 reject-axis (malformed blocks yaml.v3 accepts). The 53 reject-fixes and the 50
  accept-regressions are entangled in the same "trust yaml.v3's verdict" rule, so no
  fallback combinator (AND / OR / fallback-on-reject) is ratchet-safe. The reference
  is spec-compliant YAML 1.2; this package must be bug-for-bug Psych. This mirrors the
  sibling KEEP verdict for go-ruby-hcl2 (the Ruby value model does not fit go-cty).

What this PR contains

• No refactor, no new dependency. Production code, the emitter (dump.go,
  byte-identical to main), and the 335/402 conformance are unchanged.
• TestPsychModelInvariants (psychmodel_invariants_test.go) — a self-contained
  guard test pinning the five incompatible invariants above, so a future "wrap
  yaml.v3" attempt fails loudly. Verified to gate (fails when an invariant is
  violated).
• README "Relationship to gopkg.in/yaml.v3" documenting the KEEP verdict.

Proof

• Emit byte-exact preserved: dump.go/value.go/yaml.go diff vs main is empty;
  all TestDump* / TestOracle* golden tests green against ruby 4.0.5.
• go test -race: 100.0% coverage; gofmt/go vet clean.
• Conformance unchanged: 335/402 (83.33%), 0 regressions.

Not auto-merged (rbgo convention).