jacksonBannedFromContractsAndLoader — the foundation contracts (args, callback,
exception) and the loader infrastructure must stay free of Jackson
(com.fasterxml.jackson..); JSON marshalling belongs to value/json/parameters
(and the root Api which drives them).

All 12 architecture rules pass.

https://claude.ai/code/session_018JD9GHTJ3GNaT57iqfK3nP

Every package is its own layer with an exact mayOnlyBeAccessedByLayers accessor
set (verified against the bytecode graph), governing intra-tier edges too. All
12 architecture rules pass.

https://claude.ai/code/session_018JD9GHTJ3GNaT57iqfK3nP

…print

Root cause of the prior publish-snapshot failure: spotless:check is bound to the
verify phase, which ONLY the publish 'deploy' goal reaches — every compile/test/
package job stops before verify, so unformatted code passed all of them and only
failed at the final publish step. Fix: a new 'code-style' job (needs: startgate)
runs 'mvn spotless:check' early (fail-fast) and is now a dependency of
publish-snapshot / publish-release, so publish cannot run on unformatted code.

The same job also prints the internal package dependency graph via jdeps
(informational, continue-on-error) — the bytecode-level layering itself is already
enforced by the ArchUnit layeredArchitecture()/noPackageCycles rules in 'mvn test'.

https://claude.ai/code/session_018JD9GHTJ3GNaT57iqfK3nP

LlamaException and ModelUnavailableException already follow the unified shape
((message)/(message, cause), identity equality, typed subclass for catch-by-base),
but ModelUnavailableException had no test. Add ModelUnavailableExceptionTest
covering the constructor matrix, the LlamaException/RuntimeException type
relationship, null message, and catch-by-base. 11 exception tests green.

https://claude.ai/code/session_018JD9GHTJ3GNaT57iqfK3nP

java-llama.cpp was the only sibling repo without an org.hamcrest:hamcrest
test dependency, even though it already carried the enforcer rule banning the
legacy split Hamcrest artifacts. Add the hamcrest.version=3.0 property and the
test-scoped dependency so all four repos (BAF, sb, plugin, jllama) declare an
identical JUnit Jupiter 6.1.0 + Hamcrest 3.0 test stack. Prepares jllama for
Hamcrest matchers without forcing any test migration.

https://claude.ai/code/session_018JD9GHTJ3GNaT57iqfK3nP

…lasses

Now that Hamcrest 3.0 is on the test classpath, convert the test classes
that profit most from composite matchers (the ones whose assertions were
dominated by string-containment, map/collection membership and array-shape
checks expressed as opaque assertTrue(...) booleans):

- ResponseJsonStructureTest      assertTrue(json.contains(..)) -> assertThat(json, containsString(..)), anyOf(..)
- JsonEndpointParametersTest     same containment pattern + notNullValue()
- ModelParametersTest            containsKey/contains/array-length/same-instance ->
                                 hasKey, hasItem, containsString, arrayWithSize, sameInstance
- InferenceParametersTest        assertEquals/containsKey/startsWith/endsWith/contains ->
                                 is, hasKey, startsWith, endsWith, containsString, nullValue

Each file is converted in full (every value-assert -> assertThat) so no file
mixes idioms; assertThrows is kept (sanctioned Jupiter exception). The
containment conversions also drop now-unnecessary String derefs, so the
checks are null-safe by construction.

Verified: mvn -o test-compile clean; the two runnable pure-Java classes pass
(InferenceParametersTest 86/86, ModelParametersTest 62/62). The two
model-gated classes compile and self-skip locally; they run in CI with a model.

https://claude.ai/code/session_018JD9GHTJ3GNaT57iqfK3nP

Second pass over the suite for asserts that gain real diagnostic value
from Hamcrest matchers (collections, maps, strings, types), even where
the benefit is small. Each file is converted in full so none mixes idioms.

Matcher wins applied:
- collections      List.isEmpty()/size()  -> empty() / hasSize(n)
- maps             Map.containsKey(k)     -> hasKey(k) / not(hasKey(k))
- strings          String.contains/starts/endsWith -> containsString/startsWith/endsWith
- types            x instanceof T         -> instanceOf(T.class)
- membership       stream().anyMatch(eq)  -> hasItem(expectedPair)
- arrays           arr.length == n        -> arrayWithSize(n)
- identity/null    assertSame/assertNull  -> sameInstance / nullValue

Files: TokenLogprobTest, ModelUnavailableExceptionTest, LlamaExceptionTest,
ContentPartTest, RerankResponseParserTest, ModelMetaTest, ChatResponseTest,
ChatTranscriptTest, MultimodalMessagesTest, ChatRequestTest,
ModelParametersExtendedTest, ParameterJsonSerializerTest.

Deliberately kept as JUnit Jupiter (no clean/uglier-free Hamcrest form):
- assertThrows / fail (sanctioned exception-path asserts)
- assertEquals(expected, actual, delta) for float/double tolerance — Hamcrest
  closeTo() needs Double casts and reads worse, so converting would regress.
PairTest was left untouched: it is an equals()/hashCode() contract test
(incl. equals(null) and cross-type equals) where Hamcrest adds no value and
risks changing what is exercised.

Converting assertTrue(str.contains(..)) to assertThat(str, containsString(..))
also drops the String deref, making those checks null-safe by construction.

Verified: mvn -o test-compile clean; every runnable file green
(ModelParametersExtendedTest 140, ParameterJsonSerializerTest 35,
InferenceParametersTest-style param tests, exception/value tests, etc.);
model-gated classes compile and self-skip locally.

https://claude.ai/code/session_018JD9GHTJ3GNaT57iqfK3nP

…ckages

Bring the entire value/ and exception/ packages to 100% PIT mutation parity
and widen the pitest-maven gate from the single (stale) Pair target to the
whole net.ladenthin.llama.value.* + net.ladenthin.llama.exception.* trees.

The previous gate pointed at net.ladenthin.llama.Pair / PairTest, which the
layered-package move had relabelled to net.ladenthin.llama.value.Pair — so the
gate had silently been matching nothing. Fixed to package globs.

New tests (self-contained, no native lib / model needed):
- ChatChoiceTest, ToolCallTest, ToolDefinitionTest (new files)
- ChatMessageTest rewritten for full path coverage: plain/tool/multimodal
  ctors+factories, concatText (newline-join, image-skip, no leading newline),
  null/empty parts rejection, all three toString branches, value equals/hashCode
Killing tests added to existing classes:
- ChatResponseTest.rawJsonIsPreserved
- CompletionResultTest.rawJsonAndToStringExposeContent
- ModelMetaTest.testAsJsonReturnsBackingNode
- TokenLogprobTest.toStringIncludesTopLogprobCount
- ServerMetricsTest: predictedMs>0 boundary, asJson deref, toString content

Verified with pitest-maven 1.25.3: 145/145 mutations killed across 16 value
classes (the 2 exception classes carry no mutations); mvn test green.

https://claude.ai/code/session_018JD9GHTJ3GNaT57iqfK3nP

…ogger

Add args.* and json.TimingsLogger to the pitest gate (163/163 mutations killed).
All 10 args enums plus TimingsLogger were already at 100% via their existing unit
tests; only ContinuationMode lacked one — added ContinuationModeTest pinning
getValue() for both constants.

json.* is intentionally NOT globbed: ChatResponseParser / CompletionResponseParser
still have a few surviving mutants, and RerankResponseParser's non-array branch
returns an already-empty list (the EmptyObjectReturnVals mutant is equivalent), so
only TimingsLogger is gated from that package for now.

https://claude.ai/code/session_018JD9GHTJ3GNaT57iqfK3nP

…n/args/json)

https://claude.ai/code/session_018JD9GHTJ3GNaT57iqfK3nP

…FQNs

- json.RerankResponseParser: add testParseNode_notArray_returnsMutableEmptyList pinning
  the documented mutable-empty-list contract (matches the non-empty path, for Error Prone
  MixedMutabilityReturnType). This kills the EmptyObjectReturnVals mutant that returns an
  immutable Collections.emptyList(). Gate now includes json.RerankResponseParser
  (28 classes / 168 mutations, all killed). No production change.

- spotbugs-exclude.xml: repair the latent gate-breaker from the layered-package
  restructure — 11 <Class> entries + the OSInfo regex still used flat pre-restructure FQNs
  (net.ladenthin.llama.X) so their suppressions silently stopped matching. Updated to the
  layered packages (value/parameters/loader/callback/json). Verified spotbugs:check
  BUILD SUCCESS. Reformatted four earlier Hamcrest-converted test files to palantir style.

https://claude.ai/code/session_018JD9GHTJ3GNaT57iqfK3nP

…erage

Closes the two remaining near-100% json parsers flagged in the cross-repo
mutation-coverage expansion and adds both to the gated targetClasses list.

Production (behaviour-preserving refactors that remove equivalent mutants):
- parseChoices / parseToolCalls: collapse the "guard returns empty list /
  body builds a list" two-branch shape into a single mutable-ArrayList
  return. An empty or non-array input now falls through the loop and returns
  the same empty ArrayList, so the immutable-emptyList() empty-branch mutant
  and the redundant arr.size()==0 conditional mutants disappear, while the
  Error Prone MixedMutabilityReturnType contract is still satisfied.
- parseLogprobs: same single-return collapse.
- parseLogprobEntry: drop the emptyList()/ArrayList ternary for nested
  alternatives in favour of a single mutable list filled only when the
  nested array is present, removing the size()>0 boundary mutant.

Tests (no behaviour change; cover the previously untested typed-parse paths):
- ChatResponseParserTest: parseResponse full/multi-choice/tool-calls
  (string + object arguments)/plain/empty-choices(mutable contract)/absent/
  malformed, plus a LogCaptor test pinning the TimingsLogger.log side-effect.
- CompletionResponseParserTest: parseLogprobs post- and pre-sampling
  (top_probs vs top_logprobs), nested recursion, missing id, empty/absent
  array (mutable contract); parseCompletionResult full/limit/malformed and a
  LogCaptor timing-line test.

Verified: ChatResponseParser + CompletionResponseParser 39/39 mutations
killed; full jllama PIT gate 207/207 (100%); spotless:check clean; the two
test classes 65/65 green.

Recompile-only bump: every upstream change in the b9543..b9549 range is
absorbed inside upstream-compiled translation units (libllama + libcommon +
server-context.cpp pulled via FetchContent). Reviewed the full patch against
the project's own C++ surface — grep across src/main/cpp/ finds zero
references to any changed/added symbol (llama_context_params::ctx_other,
llama_get_ctx_other, llama-ext.h, the kv-cache mem_other/share ctor params,
n_layer_nextn, n_embd_inp_impl). Highlights:

- include/llama.h: new llama_context_params::ctx_other (shared source/target
  context); initialized in llama_context_default_params(), set by upstream
  server-context.cpp for MTP draft contexts. Project never aggregate-inits it.
- New arch GEMMA4_ASSISTANT (NextN/MTP draft head sharing the target's KV
  cache) + NEXTN_PROJ_PRE/POST tensors. A speculative-decoding/MTP feature —
  remains deferred-by-policy; loads of non-assistant GGUFs are unaffected.
- KV-cache constructors gained mem_other + share(layer_share_cb) params for
  cross-context cell sharing; all call sites updated upstream.
- common/chat.cpp LFM2 reasoning->thinking + new LFM2.5-8B-A1B.jinja template;
  handled automatically inside upstream chat.cpp.

Also fixes a pre-existing latent bug surfaced by the local build (not an
upstream change): CMakeLists.txt invoked net.ladenthin.llama.OSInfo, but the
class moved to the loader subpackage in the layered restructure, so
`cmake -B build` failed to resolve OS_NAME/OS_ARCH on hosts that don't pass
them explicitly (CI does). Fixed both --os/--arch invocations to
loader.OSInfo — same stale-FQN-after-restructure class as the earlier
spotbugs-exclude / PIT-target repairs.

Verified on Linux x86_64: cmake configure clean, libjllama.so + jllama_test
link with zero project-TU warnings, ctest 435/435 passing. Appended the
b9543..b9549 range to docs/history/llama-cpp-breaking-changes.md.