fix: Read all items when a kind holds a single key

[jsonbailey]

BEGIN_COMMIT_OVERRIDE
fix: Prevent flags from falling back to defaults when Consul holds one item
END_COMMIT_OVERRIDE

Stacked on #443. The base of this PR is jb/sdk-2995/tombstone-store-keys, not main.
Do not merge this until #443 merges. GitHub retargets this PR to main automatically when #443 lands.

Symptom

When a data kind holds exactly one key in Consul, reading all items of that kind raises:

NoMethodError: undefined method 'each' for an instance of String

The error propagates out of the store, so the whole all-items read fails rather than one item.
all_flags_state degrades to {"$flagsState":{},"$valid":false} and every flag falls back to its
default.

Root cause

ConsulFeatureStoreCore#get_all_internal read the collection with a recursive Diplomat::Kv.get.
Kv.get calls return_value(return_nil_values, transformation) and leaves return_hash at its
default false (diplomat-2.6.6/lib/diplomat/kv.rb:28, the found == :return branch). In
return_value (diplomat-2.6.6/lib/diplomat/rest_client.rb:189-203):

if @value.count == 1 && !return_hash
  @value = @value.first['Value']   # a bare String, not [{key:, value:}]
  return @value

So a recursive get returns the bare decoded value String when exactly one key matches the prefix.
.each on a String then raises. Verified against a live Consul (dev agent, diplomat 2.6.6):

keys under prefix	Kv.get(prefix, {recurse: true}, :return)	Kv.get_all(prefix, {}, :return)
0	"" (String)	[]
1	bare value String	[{key:, value:}]
2+	[{key:, value:}]	[{key:, value:}]

The fix

Read with Kv.get_all, which passes return_hash = true and so always returns key/value pairs,
and which returns [] for a 404 when given not_found = :return
(diplomat-2.6.6/lib/diplomat/kv.rb:106-132). get_all sets :recurse itself. It returns full
keys, so the existing prefix-stripping is unchanged. The results == "" guard for the empty case
becomes dead and is removed, because get_all gives [] instead of "".

Why this matters for #443

The .each line is not changed by #443, so the bug is pre-existing. But it defeats #443's own fix
at n=1: a store whose only features row is a keyless tombstone — a single-flag project, or every
flag deleted — still raises instead of reading the tombstone and filtering it. #443's new spec does
not catch this because it seeds a live item plus a tombstone, which is two keys.

Test evidence

Two specs added to the shared persistent_feature_store examples in
spec/feature_store_spec_base.rb:

• can read all items when a kind holds a single item
• can read all items when the single item is a tombstone with no key

Before (on jb/sdk-2995/tombstone-store-keys, 2 specs x 4 permutations):

8 examples, 8 failures

NoMethodError:
  undefined method 'each' for an instance of String
# ./lib/ldclient-rb/impl/integrations/consul_impl.rb:73:in 'ConsulFeatureStoreCore#get_all_internal'
# ./lib/ldclient-rb/integrations/util/store_wrapper.rb:98:in 'CachingStoreWrapper#all'

After, the Consul suite is green:

92 examples, 0 failures

Full suite against live Redis, Consul and DynamoDB (LD_SKIP_DATABASE_TESTS=0):

1394 examples, 0 failures

bundle exec rubocop: 187 files inspected, no offenses detected.

Note on spec placement

The specs go in persistent_feature_store rather than any_feature_store, so they also cover
Redis and DynamoDB. Both already pass there (16 examples), which confirms the defect is specific to
the Consul client.

They deliberately read through a second store instance. CachingStoreWrapper#init warms the
all-items cache (store_wrapper.rb:60-74) and caching is on by default at a 15s TTL, so a
single-instance all after init is served from cache and never calls get_all_internal — it
passes even on the broken code. An any_feature_store placement would therefore not have caught
this, and would not work for the in-memory store, which has no shared backing for a second
instance and no write_raw_item.

---

Note

Overview
Fixes Consul feature store all reads when a data kind has zero or one key under its prefix. Listing used recursive Diplomat::Kv.get, which returns a bare value String (or "") instead of key/value pairs when only one match exists—each then raised and all_flags_state could degrade to invalid defaults.

get_all_internal now uses Diplomat::Kv.get_all, which always returns an enumerable list (including [] when nothing matches), so prefix stripping and tombstone handling stay the same without the old empty-string guard.

Shared persistent_feature_store examples add coverage for single-item, empty-kind, and single keyless-tombstone cases, using a second store instance so reads hit the database rather than the warmed all-items cache.

^{Reviewed by Cursor Bugbot for commit 2734305. Bugbot is set up for automated code reviews on this repo. Configure here.}

[jsonbailey]

Checked get_all availability across every diplomat version we allow, since the switch depends on it.

The pin is ~> 2.6 (launchdarkly-server-sdk.gemspec:31, and contract-tests/Gemfile:17), so the range is >= 2.6.0, < 3.0 — in practice 2.6.0 through 2.6.6, as 2.6.6 is the newest 2.x. All seven are identical on every property this fix relies on:

	def get_all	passes return_hash = true	[] on 404	count == 1 && !return_hash guard
2.6.0 – 2.6.6 (all 7)	yes	yes	yes	yes
upstream master	yes	yes	yes	yes

Two notes:

• get_all was added in 2.3.0 (absent in 2.0–2.2, unchanged since), so the pin has three minor versions of headroom. Loosening it to ~> 2.3 would still resolve.
• master still has get omitting the third argument and the collapse guard intact, so this is current upstream behavior rather than a quirk of the pinned release — the n=1 collapse is not something already fixed upstream that we would be working around.

Out of scope for this PR, but worth recording: diplomat is only a development dependency, so Consul users supply their own. Someone on 2.0–2.2 would hit NoMethodError on get_all. That exposure is not new — the gemspec has never constrained a user's diplomat at runtime, and the same applies to every other Diplomat::Kv call in the file — but a documented minimum or a runtime dependency would close it.