perf: retain full Jet channels after preservation analysis
Record T4 as a nonadopted request-specific derivative-channel candidate. No product, private helper, public API, fixture or test-registration changes are required; the production sources and archives remain identical to T3. Original-source preservation tests show that Compose with a high nonfinite third coefficient and zero direction retains finite value/gradient/Hessian but produces mixed-second NaN, IsFinite failure and INVALID FE/MXCSR state. An overflowing high reciprocal coefficient reaches the same boundary. Negative scalar times zero also preserves signed-zero directional/mixed channels, so replacing unused fields with initialized +0 changes behavior. The omission candidate was rejected without a validation-replication framework. Copied-source instrumentation in three valid representative small HHT tests recorded API entry attempts before validation: none 42, explicit-zero 78, nonzero 47. These are scoped test-entry observations, not successful or accepted evaluation counts or official-model frequency estimates. Normal Release request-mode timings are descriptive auxiliary costs, not channel cost attribution or a before/after speedup. No official model ran in T4. Separate T4 original/current supplemental comparisons passed in six routes; six explicit omission-mutant comparisons returned the expected exit 1. Unchanged T3 build/default/related evidence is linked by exact source/archive hashes. Existing physics failures and full T6 verification remain unchanged. Artifact root: C:/git/FESADev/.build/jet-product-20261021-01/t4 T4-summary.json SHA256: c348a62d6c686c667286ba8ee16f310134b13c779132fe839765def89af568c0 T4-summary.md SHA256: ecfea9771dd8de396d86394262f393a3da9c480b8a80c500a9545a383b80bb15 commands-manifest.json SHA256: 4cace21f3b49932f52d9cf1bd9dfdc2ad7d9ccf40657da8849c678dd24e0a538
This commit is contained in:
KOKO\Mimi