Fix: Improve anchor program deploy reliability with resumable buffers

[swaroop-osec]

Summary

• Makes anchor program deploy and anchor program upgrade resumable across retries and reruns using persistent per-program buffer keypairs.
• Improves priority-fee selection by querying contention-aware recent fees for the accounts being write-locked.
• Honor --skip-preflight consistently across buffer writes, deploys, upgrades.

Closes #4481

[vercel]

@swaroop-osec is attempting to deploy a commit to the Solana Foundation Team on Vercel.

A member of the Team first needs to authorize it.

[greptile-apps]

Greptile Summary

This PR overhauls anchor program deploy and anchor program upgrade to be resumable across retries and reruns: it introduces per-program persistent buffer keypairs stored under target/deploy/, a fetch_buffer_program_data helper for diff-only resume (re-sending only un-landed chunks), an outer retry loop in both deploy and upgrade paths, and consistent --skip-preflight and priority-fee propagation through all transaction types.

• Persistent buffer + diff-only resume: ensure_buffer_keypair_arg pins a per-program keypair; prepare_write_messages skips already-matching chunks on each attempt, enabling efficient recovery from partial failures.
• Contention-aware fees: get_recommended_micro_lamport_fee now accepts write_locked_accounts so the fee query is scoped to transactions that write-locked the same program/buffer accounts, replacing the global-median approach.
• Skip-preflight propagation: deploy_program, upgrade_program, close_undersized_buffer, and all write batches now honour the caller's skip_preflight flag rather than always running preflight.

Confidence Score: 3/5

The deploy/upgrade paths work correctly for the common case, but the persistent-buffer resume logic has a correctness gap when the user shrinks the program binary after a failed full-buffer-write attempt.

The stale-tail-bytes issue can silently deploy a corrupted program binary or fail ELF verification in a scenario the PR was specifically designed to handle — fully-written buffer, failed upgrade tx, then a smaller binary on retry.

cli/src/program.rs deserves a close second look, particularly the oversized-buffer case in both program_deploy and program_upgrade, and the commitment level passed to fetch_buffer_program_data.

Important Files Changed

Filename	Overview
cli/src/lib.rs	Adds write_locked_accounts param to fee-query helpers for contention-aware priority fees; aligns DEFAULT_MAX_SIGN_ATTEMPTS to 5 (Agave default); moves --buffer injection out of this helper and into per-program callers; removes the global add_recommended_deployment_solana_args call from the workspace deploy() wrapper.
cli/src/program.rs	Large refactor introducing resumable buffer logic: persistent per-program keypairs, diff-only resume, close_undersized_buffer, outer retry loop, and skip_preflight propagation. Key issues: stale tail bytes when a fully-written larger binary's upgrade fails and the user then shrinks the binary; fetch_buffer_program_data uses potentially-finalized RPC commitment; CreateBuffer tx lacks a compute unit price.

_{Reviews (1): Last reviewed commit: "fix(cli): ensure deploy directory is val..." | Re-trigger Greptile}