CARLA sampled value bounds

Derivative-free sampled cover bounds

Sampled lane covers need value/error enclosures, not derivatives. The private lane_value_bounds.mojo path evaluates the same POLY3/PARAM_POLY3 expression with _ValueJet. Direct full-Jet APIs remain available for comparison. Four closed quarter domains, shared endpoints, work debits, finite admission and optional failure behavior are unchanged.

Proof scope

For the reviewed _JetExpression primitives, constants and variables give the same value intervals and zero scalar error. Negation, addition, subtraction, multiplication, division and rounded_value() compute value/error using only operand value/error fields. Derivative calculations populate separate fields; omitting them does not change those operations. Derivative-free fields stay whole and are never a derivative proof. The square-root primitive is excluded because its zero shortcut inspects derivative fields.

The sampled graph retains every sample-index, clamp, finite, sign, reduction and signed-axis guard. Stored constants and shared trig/Horner expressions do not change. Its conditioned blend keeps the original weighted scalar-error expression and the algebraically equal local ideal-value expression, with the same value intersection. Only the unused derivative intersections disappear. Lane record, width, offset and elevation traversal and final center expressions retain their original operation grouping.

tools/carla_lane_oracle/check_sampled_values.py checks all eight copied helpers, four value dispatch helpers, fourteen shared generic/full-Jet dispatch helpers, actual unique aliases/decorators, exact imports and the reviewed primitive graph. The complete normalized AST of all three reference modules is conservatively bound, including imports, aliases, union setup and skipped dispatch regions.

Any dependency edit requires renewed source review. It rejects derivative field/flag reads, square roots, missing/extra helpers and source divergence. It uses explicit failures under optimized Python. It does not execute Mojo or regenerate production source. Twenty-five compiler-free controls in tools/test_sampled_values.py run through make test-tools, including stale full-Jet changes, arithmetic/guard/type mutations and optimized-Python refusal.

Run the source check explicitly with:

python3 -B tools/carla_lane_oracle/check_sampled_values.py

The native equivalence test checks all value/error/rounded-bound words on 1,940 Town quarter domains and all 24 interval endpoint words in each of 485 covers. It also covers scales, signed-zero/atan branches, budgets and optional missing-record failure. Finite controls supplement the source argument; they do not establish arbitrary-domain correctness by sampling. Original cover, Town and sampled-axis suites also pass on the experimental source.

Measured scope and remaining regression

A matched portable x86-64-v3, default-O3, Werror, single-thread experiment used one warmup and eight balanced comparisons with 144 consumed route queries per run. Median construction was 48.552 ms versus 60.097 ms for the prior qualified lane candidate; the median paired ratio was 0.79375, a 20.6% reduction. All 36 returned station, center, certificate and work-counter records were identical.

Queries measured 5.419 ms versus 5.807 ms, but unchanged query work means this is not evidence of an algorithmic query reduction. Against main in the same window, median paired slowdowns were still 32.46x construction and 38.71x query. This remains a substantial performance regression. The experiment does not claim final composed-source, full-suite, coverage or general performance acceptance; those gates remain separate.