Configured native lint selection — 2026-09-08¶

The local make lint-c driver at cc022245079f1b7da4005a39bcefc759e2d6ef3d overwrites Meson's database with unfiltered ninja -t compdb, then selects sources using core/src/**/*.c, core/src/**/*.cpp and core/tools/*.c. The first two Git pathspecs omit engine-root files; the third omits C++ tools. Tests are absent. Conversely, inactive ARM and CUDA files are selected without configured commands, producing unrelated inferred-command diagnostics.

The retained CPU profile contained 1,215 native Meson command records. Ninja's unfiltered export expanded that to 1,411 records including generation/link commands. The Make source list selected 216 files: 99 had no configured entry, while 157 configured tracked sources were omitted. These are profile-specific counts, not the canonical hosted GCC14 ratchet inventory. The full lint attempt was interrupted after confirming driver defects, and is not a passing receipt.

The first observed clang parser failure was GCC -flto=4. The existing hosted lint job already avoids it with a separate -Db_lto=false profile. A private analyzer database instead maps positive numeric -flto=N to -flto, preserving the original build's configuration and byte-identical input database. Existing clang spellings and unknown values are not normalized into success.

The implementation follows ADR-1142's configured-source coverage policy. No new ADR is needed: this repairs the local selector without changing the ratchet baselines, diagnostic thresholds or remote job scope. The existing ratchet loader selects source paths from its database, then clang-tidy reads all matching commands; selecting a source once must not discard distinct macro/include/language variants and optional output metadata. Local lint also retains tracked vendored sources. Generated/untracked sources and inactive tracked sources are recorded separately; the ratchet's generated-source policy remains independent.

Alternatives: adding globs leaves inactive files eligible and misses future extensions; selecting one command per filename loses meaningful test variants; reconfiguring the user's build changes its requested profile. The selected approach keeps the native database and filters a private copy against Git's tracked native paths, preserving all matching entries. Both analyzers run so one failure does not hide the other's findings.

Validation uses scratch Git repositories and real subprocess boundaries with analyzer stubs, including the actual Make target. Controls cover root engine, C++ tools, tests, tracked vendor and CUDA sources; inactive ARM and generated inputs; distinct macro variants; original database preservation; command and arguments schema; numeric-only LTO adaptation; missing/invalid inputs; and independent analyzer failures. These are driver regressions, not native runtime or numerical tests. Netflix golden assertions are untouched.

Retained machine evidence: .workingdir2/cache/rc1-full-lint-20260908-cc022245/ contains both interrupted full-Make attempts, tool/image hashes, command logs, Meson/Ninja/analyzer help, and the scope inventory. The original 134/134 CPU test receipt is under .workingdir2/evidence/2026-09-08-rc1-full-gates-cc022245/; it does not establish golden, sanitizer or GPU acceptance.

The LLVM compilation database specification permits multiple commands for a source and an optional output field that distinguishes processing modes. The analyzer copy retains this metadata and validates its type alongside the command schema.

Migration check: the actual 1,411-entry export contains all 1,215 native commands unchanged plus 196 non-native extras. The 48–51 DNN variants are genuine test configurations already in the native database. Ten extras are phony entries with empty commands; strict schema validation rejects them. In a disposable copy, meson setup --reconfigure BUILD SOURCE/core restored 1,215 entries without changing the JSON build-options inventory. Make runs this no-override reconfigure before building and analyzing. The helper never rewrites the input database itself. A no-op-build fixture verifies old phony entries are repaired; ordinary reconfiguration does not discard genuine source-command variants.