perf(planner): prefer narrower roots within payload slack

[markosg04]

Summary

The scalar planner currently chooses the smallest proof payload even when a negligible payload difference requires a wider root commitment. This adds an opt-in payload budget: find the minimum-payload candidates, admit candidates within ceil(min_payload * (1000 + slack_permille) / 1000), then choose the smallest root inner rank and payload. Recursive configurations carry the same budget into direct fallback.

The Akita default remains zero; Jolt's adapters use 10 (1%). The parameter is included in generated-catalog identity, and every selected candidate still passes the existing SIS, fold-bound, setup-envelope, and schedule checks.

Testing

• Ran tests for modified crates
• cargo fmt --all --check passes
• Relevant cargo clippy mode passes

The full local nextest suite passes (1,202 tests), schedule regeneration is clean, and all three required Clippy configurations pass with warnings denied.

Security Considerations

• Verifier acceptance behavior is unchanged, or the intended change is specified.
• Verifier-reachable malformed inputs return typed errors instead of panicking.
• Transcript labels, challenge order, and domain separation are unchanged, or the intended change is specified.
• Serialization changes preserve canonical decoding and bounded untrusted input handling.
• New dependencies, Git dependencies, or CI actions are justified and pass supply-chain policy.
• Unsafe code is unchanged, or each new unsafe block has a local safety argument.

This changes schedule selection and catalog identity, not verifier equations or Fiat-Shamir ordering.

Breaking Changes

SelectionPolicyId::MinFirstDirectSetupThenPayloadWithinSupportedEnvelope now carries its direct-fallback payload slack. Generated catalogs using this policy need regeneration.

[markosg04]

Closing per review feedback (thread with @quangvdao): the payload-slack selection policy comes out of the stack entirely. #345 now reverts the planner, config policy surface, and catalog identity to their main state and keeps only the prover optimizations; the Jolt-side adapter and catalogs are being regenerated at the default min-payload policy.

One data point worth keeping for the planner rework: at the 2^28 D128 shape, the rank-3 / P=2^21 schedule proves 15.6% faster with ~4 GiB lower peak RSS and an identical proof payload to the selected rank-4 / P=2^18 schedule — that case is an equal-payload tie-break, not a payload trade. Happy to pick that up as a planner-native tie-break or via the adaptive-dimension candidate sets once #356/#355 settle.