perf: adopt ring-go perf/hash-cache off-chain signing optimization

[oten91]

What

Adopts ring-go's perf/hash-cache off-chain signing optimization as an opt-in, named path alongside the existing deterministic Sign — mirroring ring-go's own Sign vs SignWithContext split.

Supersedes the perf/signer-context-cache approach (which overloaded Sign with off-chain-only behavior).

Changes

• signer.go
  • Sign — unchanged behavior: deterministic, consensus-safe (Ring.Sign). Existing callers keep identical semantics.
  • SignOffChain / SignOffChainWithRing — new opt-in off-chain fast path (Ring.SignWithContext). Caches one secret-free SignerContext per ring (sync.Map), reused across signs; SkipSelfCheck enabled (safe — off-chain, always signed with the scalar the context was built from). Off-chain only.
  • ClearSignerContextCache for session rollover; private-key scalar decoded once at construction.
• signer_test.go — round-trip verify for both paths, tampered-message reject, cache reuse + clear, asserts Sign never touches the cache.
• deps — ring-go → perf/hash-cache (0830a5c); go-dleq now direct; go-ethereum → v1.17.4 (CVEs, cgo-only); x/crypto/net bumps.
• docs — docs/migration-ring-go-hash-cache.md for PATH.

Downstream (PATH etc.)

• On signer-context-cache, Sign was the cached path. Here Sign is deterministic again — swap Sign→SignOffChain, SignWithRing→SignOffChainWithRing to keep the speedup. Leaving Sign as-is stays correct, just slower. See the migration doc.
• Requires the Go 1.26 toolchain (propagates from the SDK).

Verification

Build ✅ · vet ✅ · golangci-lint 0 issues ✅ · tests pass (both paths verify) ✅

⚠️ Draft — do not merge yet

ring-go is pinned to a branch commit, not a release tag. Keep off main until go-ring promotes hash-cache to master/tags it, and until the gateway reports off-chain before/after numbers.

🤖 Generated with Claude Code

[oten91]

✅ Canary validation (PATH, draft PR #517)

PATH adopted this branch and validated on canary (beta-4, node-matched, 82 min, rc=0, live HTTP + WebSocket + hedge traffic):

• Adoption: 1-line swap SignWithRing → SignOffChainWithRing; SDK-side SkipSelfCheck worked transparently. Matched PATH's existing per-session ring caching.
• Micro-benchmark: hash-cache vs precompute — −17.2% allocs / −16.2% bytes at ring 32, scaling to −12% at ring 2. In the predicted ~15–18% window.
• Canary: PATH is signing-CPU-bound (~54% of pod CPU in SignWithContext → signInternal; verification only ~4.5%). Signing CPU materially down — secp256k1 FieldVal.SquareVal flat 31.9% → 23.6%; signing CPU share dropped and deepened as the context cache warmed (precompute amortization confirmed). No regressions, signatures verify at the relay miner, rc=0 throughout.
• Backend: pure-Go decred secp256k1 (no CGO) as expected — win is fewer field ops.

Follow-up applied

• Documented the unbounded signerContextCache: consumers MUST call ClearSignerContextCache() on session rollover or it leaks slowly (437b446). Decision: keep eviction consumer-managed (doc-only), not an internal LRU/TTL — PATH already wires rollover eviction.

Still gated on

• ⛔ ring-go must cut a tag for hash-cache. Then: repin SDK branch-commit → tag, un-draft, merge, tag SDK release, PATH re-pins. This PR stays draft until then.