This document is the citable statement of launchbound's safety decision
rule. It is stated here and in the README. Changing
this rule requires explicit human sign-off — it is the product.
For each candidate configuration, launchbound invokes:
cargo reconverge check --strict --message-format json --cc <target>
--strictis mandatory, not optional. The findings this product exists to catch are warning-tier byreconverge's design: a launch contract is a declaration, not a proof, so launch-shape-dependent findings never promote to a gate on their own andcargo reconverge checkexits 0 on them at default confidence. An autotuner does not have that excuse — it is choosing the launch configuration itself, so it can supply the contractreconvergelacks and turn the warning into a decidable answer.--cc <X.Y>is supplied per target because shared-memory capacity context (RC004) depends on it, and shared-memory layout is a tuning dimension. A safety verdict obtained at one--ccdoes not transfer to another (T4 is--cc 7.5; A10G is--cc 8.6).- Exit codes:
0clean,1findings at deny/confirmed,2tool error. A2is a hard stop for that candidate, never a pass by omission.
launchbound parses reconverge's findings.v1 output and records, per
candidate: rule ID, confidence, and the source span. That record is what the
rejection report prints.
For a candidate configuration evaluated at its own launch configuration:
- Disqualified —
reconvergereports a barrier or masked-collective finding whose divergence source is launch-shape-dependent. - Admitted with a recorded caveat — a finding exists but is launch-shape-independent. The caveat appears in the report.
- Clean — otherwise.
A disqualified configuration is never selected, never ranked, and always listed in the rejection report with its rule ID and source span — especially when it was faster than the configuration chosen.
--allow-unsafe exists for users who know their launch contract differs from
the declared one. It requires an explicit reason string, the reason is
recorded verbatim in the report, and it is never the default. A missing
reason string is a usage error, not a warning.
The gate inherits reconverge's documented limits wholesale: summary-based
interprocedural analysis, reducible control flow only, non-literal masks not
evaluable, and data races entirely out of scope. A clean gate is not a
proof of correctness. It means: at this launch configuration, the analyzer
found no barrier or masked-collective divergence hazard within its documented
coverage.
The Metal backend has no gate at all — there is no reconverge
equivalent for MSL and this project does not build one. Every Metal report
header carries that notice, and a test asserts the notice cannot be omitted.