fix(llama-cpp): constrain strict LFM arguments

backend/cpp/llama-cpp/tests/README-native-strict.md

@@ -0,0 +1,63 @@
+# Local native LFM schema regression
+
+`patches/0003-lfm-strict-schema.patch` targets llama.cpp
+`c0fd78626405c6d9b934efef0ec7291fcc20e3d6`. It preserves the native strict
+flag through tool parsing and serialization, then selects a schema-aware
+Python-style argument parser for strict tools. Non-strict tools retain the
+legacy builder. This patch is carried by LocalAI; no upstream submission is
+required.
+
+The supported subset covers Sound's object properties, required fields,
+optional omission, string enums, string lengths, bounded integer/decimal
+values and its musical-meter pattern. Unsupported schema constructs fail
+explicitly rather than silently dropping constraints. Strict generation uses
+canonical property order and decimal notation. The parser also accepts valid
+reordered arguments, equivalent escaped enum strings and in-range exponent
+notation, and validates their decoded values. LFM final responses must have a
+complete native envelope before tools are published, including non-strict calls.
+Partial streaming and non-strict argument recognition remain unchanged.
+
+## CPU-only checks
+
+Use an isolated extraction of the exact pinned source. Apply the patch with
+`git apply --check` followed by `git apply` inside an initialized isolated
+source repository. The runner neither downloads nor patches source. Do not
+run Git patch checks from an uninitialized directory inside another repo:
+Git can resolve that enclosing repository instead of the intended source.
+
+From the LocalAI root:
+
+```sh
+backend/cpp/llama-cpp/tests/run-native-strict.sh /absolute/pinned-source /absolute/scratch-build
+```
+
+The runner builds the actual `llama-common` and `llama` CPU libraries and runs
+all three CTest executables. It requires no model, inference, GPU or gRPC server.
+Rebuild dependent objects after changes to the native tool struct. The
+harness names deliberately do not match the standard-library-only backend
+runner's `*_test.cpp` discovery pattern.
+
+`native-strict` checks strict true/false/absent round trips and legacy
+three-field aggregate initialization. `native-schema` uses the committed
+`fixtures/sound-schema.json`, derived from LocalAI's Sound tool definition,
+and exercises generated grammar recognition, complete PEG parsing and mapped
+tool calls separately. This fixture is covered by the repository's MIT
+license. It includes all 504 existing keyscale values.
+
+The baseline accepts 30 invalid cases in the original 44-case matrix. The
+expanded patched matrix currently has 1,092 cases, including all enum values
+under both parallel settings. `native-edge` adds 639 checks for semantic edges,
+required/auto/none modes, strict true/false/absent and bytewise final truncation,
+plus six JSON-comparison controls. String repetitions use bounded chunks because
+the native GBNF reader treats repetition bounds above 2,000 as unbounded.
+These are offline parser checks, not tokenized sampling or model benchmarks.
+
+## Maintenance and remaining integration
+
+Keep the native patch tied to the exact pin above. Rebase and rerun transport,
+grammar and compatibility checks whenever that pin changes. Do not assume
+an upstream update fixes this behavior without running the actual LFM path.
+Production patch preparation, full schema compatibility review, additional
+regressions, performance comparison, duplicate-key defense and CI integration
+are separate acceptance steps. Do not treat this harness alone as a deployed
+or complete repair. No inference-throughput improvement is claimed.