Harden autonomous AgentPool and prepare v4.4 mainnet candidate

[Ohjjam]

What changed

• closed GPT-reviewed finance failures: unregistered-worker dead ends, validator no-quorum slashing, proof-window expiry, and mutable same-job replans
• prevented epoch-reservation stacking and made dynamic Issue budget, candidate, and operator-group admission lifetime-bounded
• made runtime profiles append-only and Work Power additive at a stable governance snapshot
• enforced five voters, three operator groups, 30% total Work Power quorum, and exact two-thirds cast support for MATURE Issue and release decisions
• excluded the TRANSITION proposer group, rejected duplicate validator groups, and used a conservative lower median
• bound all seven release gates to real evidence-file hashes, exact source identity, deployment journal identity, constructors, creation bytecode, and runtime bytecode
• added one exact v4.4 Base Sepolia/mainnet deployment engine with testnet-only acknowledgement and no gate weakening
• added a 90-day public reliability campaign, append-only observation intake, deployment-registered observer attestations, and RPC semantic verification
• pinned canonical reliability regeneration to an exact block while separately enforcing current wall-clock and chain-head freshness
• bound reverted cap and Issue-replay evidence to exact calldata, prior-block state, admission preconditions, and custom errors
• required every release entrypoint dependency and the canonical reliability policy to exist in the committed Git tree
• recorded three maintainer-directed GPT review tracks; all returned NO_CONFIRMED_BLOCKER on the final tree while independent review remains a separate gate

Verification

• Solidity compile: 71 contracts
• exact v4.4 mainnet graph rehearsal: 276 transactions / 33 checks / 24 objectives
• mainnet finance and state rehearsal: 508 transactions / 415 checks
• public autonomous-economy compatibility rehearsal: 731 transactions
• focused v4.4 gate regression suite: 49/49
• repository tests: 165/165
• Slither 0.11.6 over 14 deployed implementations: reviewed baseline reproduced; no new High or Medium findings
• reproducible source evidence SHA-256: c944b297d31a0eb16af350338ee7f83794ec36066815546b7d24ca0d9e94cc71
• exact evidence commit: 2697463fec635beb3e0d970d167228d3e59175c7

Review entrypoints

• audits/V44_GPT_COLLABORATIVE_REVIEW.md
• audits/V44_GPT_REVIEW_PACKET.md
• audits/V44_SLITHER_TRIAGE.md
• MAINNET_V44_RUNBOOK.md

Deployment boundary

This remains a draft, fail-closed mainnet candidate. No Base mainnet transaction was sent.

The v4.4 Base Sepolia deployment and its 90-day live campaign have not happened yet. Mainnet remains blocked until the exact public campaign supplies at least 180 RPC-verified transactions, five contributing agents, three operator groups, two registered independent observer attestations, fresh indexing, all required adversarial paths, and zero unresolved critical invariant incidents. Independent security review, independent economic review, validator operational-independence evidence, the actual deployer's legal assessment, and name/symbol clearance also remain required.

[Ohjjam]

Superseded by #24, which contains the complete reviewed and verified v4.4 candidate merged into main.