feat: KV v2 support + generic verbs (opt-in, non-breaking)

[kurok]

What

Lets the simple client work against KV v2 mounts and any other Vault backend, lazily and transparently — callers don't need to know the engine version — while staying non-breaking for current users.

• Opt-in, default off. api.kv.autoDetect (default false) → with it and api.engines unset, read/list/write are byte-for-byte the old behavior and no sys/internal/ui/mounts call is made.
• Lazy auto-detect. When autoDetect:true, each mount's KV version is detected on first use via GET sys/internal/ui/mounts/<path> (readable by least-privilege tokens), then paths are rewritten and responses unwrapped.
• api.engines override. Static { secret: 2, legacy: 1 } map; listed mounts skip detection. With engines set but autoDetect:false, unlisted mounts passthrough (no detection call) — so it's a working fallback for Vaults that deny the detection endpoint.

Public API

Method	Behavior
read/list/write/delete	version-aware when enabled; passthrough otherwise
update(path, data)	PATCH with application/merge-patch+json + {data} envelope (KV v2)
request(method, path, data)	raw, literal path, no rewrite/unwrap, parsed body (transit/pki/database/sys/…)
deleteVersions/undeleteVersions/destroyVersions/readMetadata/deleteMetadata	KV v2-only; throw UnsupportedOperationError on v1/non-kv
Lease.getMetadata()	KV v2 version metadata; undefined on v1 (additive)

Design

• src/kvTransform.js (new, pure, no I/O) — rewritePath / normalizeResponse encode the full op × {v1,v2} table.
• src/MountResolver.js (new) — injected detectFn, cache by canonical mount path, longest-prefix lookup, in-flight Promise dedupe, engines bypass, actionable VaultError on detection failure.
• VaultApiClient.makeRequest — minimal change: only set the default Content-Type when the caller didn't supply one (enables merge-patch).
• Lease — optional metadata (additive); errors — add UnsupportedOperationError.

How this was built & reviewed

Implemented via a multi-agent workflow (TDD build → 3-reviewer panel: architecture/consistency, correctness, acceptance → fix), then independently validated. Bugs caught and fixed before this PR:

• Trailing slash stripped in passthrough — list('secret/') hit secret instead of secret/, violating the byte-for-byte guarantee. Fixed: version!==2 now uses the caller's literal path. Regression tests added.
• MountResolver concurrency collision — concurrent first-touch of two distinct multi-segment mounts sharing a first segment (team/kvA, team/kvB) collapsed into one detection and mis-classified the second. Fixed: followers verify the resolved canonical mount actually prefixes their path; else a fresh detection runs. Regression test added.
• autoDetect:false + engines set — an unlisted mount still attempted detection. Fixed: detection only runs when autoDetect is on; engines-listed mounts resolve via override, everything else passthrough. Regression test added.
• request() falsy body — data || null dropped 0/''/false; now data === undefined ? null : data.
• Double-rewrite hazard (secret/data/foo + autoDetect) documented in the README.

Tests

• npm run test:unit — 237 passing (new kvTransform + MountResolver suites are exhaustive; VaultClient/conformance cover every acceptance criterion)
• npm run lint — clean

Scope notes

Within the agreed file list, plus a 9-line addition to test/errors.test.mjs covering the new error class. No version bump / CHANGELOG (handled at release). Out of scope: auth backends, typed transit/database helpers, index.d.ts.