deploy_pda_seed's first paragraph just restated its parameter names;
kept only the domain-separation rationale. chained_call_forwarder.rs's
doc comment was floating over an unrelated type alias instead of being
a module-level //! doc.
Removed explanatory comments that restated context better left to commit
history/PR description across the deploy-dispatch and program-storage code.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Introduces a new loader_program/loader_core crate pair implementing a
Deploy instruction that claims a program's ProgramData PDA account
(image_id, segment_number, update_auth, elf_segment), unifying
deployment with ordinary PublicTransaction dispatch instead of the
separate ProgramDeploymentTransaction path.
Measured against every real program in this repo, computing a
program's image_id inside the zkVM costs ~1,400-1,500 cycles per byte
of deployed bytecode, pushing real deployments to 500M-900M cycles
against the 32M public-execution cap (vs. ~27ms natively, since
ProgramDeploymentTransaction's equivalent check runs as a plain host
function today). To keep the unified dispatch path viable, Deploy is
special-cased in from_public_transaction: calls targeting the reserved
RESERVED_DEPLOYMENT_PROGRAM_ACCOUNT_ID run loader_core::execute_deploy
natively instead of through the interpreted guest executor, wrapped in
catch_unwind since the shared execute_deploy logic validates via
assert!/expect() like every other guest program, relying on that
boundary instead of the zkVM's own panic-to-Result conversion.
The loader guest binary is kept buildable and covered by a test that
runs it for real and asserts its output matches the native path
exactly, so the two can't silently drift apart.
Marvin Jones