starknet_transaction_prover: classify mod-builtin prover panics pending typed errors

[avi-starkware]

Summary

Temporarily classify a prover-task panic as UnsupportedBuiltins when the PIE used add_mod or mul_mod components absent from the recursive verifier's component set. Map that error to JSON-RPC 1002 (Unsupported builtin).

The prover still runs normally. Successful proofs are unchanged; returned errors, cancellations, and panics without unsupported-builtin usage remain internal errors.

This is a heuristic: an unrelated panic in a transaction using an unsupported mod builtin will also produce 1002. A TODO at the classification branch calls for a typed unsupported-component error from privacy-prove, replacing this inference with direct error handling.

• Use JoinError::is_panic() and positive PIE builtin counts. No panic capture module, thread-local storage, or source-location matching.
• Leave the existing observability panic hook unchanged and preserve the original documentation in prover.rs.
• Pin the spec to feat(proving-api): add UNSUPPORTED_BUILTIN error to starknet_proveTransaction starknet-specs#411 at 8c315cb2d3bde23379d92b415c2371ce68b51643, report spec version 0.10.4, and validate the new RPC error against the schema.
• Test panic/cancellation classification with and without mod-builtin usage, display formatting, and RPC mapping.
• Keep real-proving tests for mul-only and add-plus-mul transactions alongside the multicall success test. The tests assert builtin usage and the returned error, not a panic source location.

Stacked directly on #15124 over main-v0.14.4. This replaces the approach that depended on #15125's panic-capture machinery.

Verification

• Proving-enabled library suite: 153 passed, 6 ignored.
• Non-proving library suite: 140 passed, 5 ignored; one doc test passed.
• Real-proving flows: 3 passed.
• Clippy with -D warnings passed in both configurations, as did formatting.
• Original prover.rs docstrings match the base verbatim. The observability hook and its tests also match the base.

cargo +nightly-2026-01-15 test --locked -p starknet_transaction_prover --features stwo_proving --release --lib --quiet
cargo +nightly-2026-01-15 test --locked --release -p starknet_os_flow_tests --features stwo_proving prove_ -- --ignored --test-threads=1
cargo +nightly-2026-01-15 clippy --locked --release -p starknet_transaction_prover -p starknet_os_flow_tests --features stwo_proving --all-targets -- -D warnings
cargo test --locked -p starknet_transaction_prover --quiet
cargo clippy --locked -p starknet_transaction_prover -p starknet_os_flow_tests --all-targets -- -D warnings
scripts/rust_fmt.sh --check

[reviewable-StarkWare]

This change is 

[cursor]

PR Summary

Medium Risk
Changes proving error classification and public RPC semantics (new 1002) using a heuristic that may mislabel unrelated panics when mod builtins were used; successful proofs are unchanged.

Overview
When proving fails on a blocking task panic and the Cairo PIE used add_mod / mul_mod builtins that are absent from PRIVACY_TRANSACTION_COMPONENTS, the prover now surfaces ProvingError::UnsupportedBuiltins instead of a generic join failure. That maps to JSON-RPC 1002 (Unsupported builtin) with builtin names and instance counts in data; other proving failures stay internal errors.

The proving path scans the PIE before privacy_recursive_prove, wires privacy-circuit-verify-v2 under the stwo_proving feature, bumps the pinned proving API to 0.10.4, and documents the limitation. CI runs all ignored prove_* flow tests via a local stwo_proving feature on starknet_os_flow_tests; integration tests add try_prove and prove_tx_using_mod_builtins_fails to pin the failure mode.

^{Reviewed by Cursor Bugbot for commit c23cf02. Bugbot is set up for automated code reviews on this repo. Configure here.}