perf(aot): reduce preflight latency with cached full-coverage artifacts

[hero78119]

Problem

Preflight AOT coverage was derived from a bounded execution sample and a fixed set of hot roots. Late control-flow targets could therefore remain uncompiled, while each process repeated training and native compilation even when the program and AOT ABI were unchanged.

Design Rationale

A lightweight coverage tracer now follows the complete training workload and records only control-flow roots. This provides whole-workload basic-block coverage without constructing full preflight RAM and shard state during training.

Compiled artifacts are persisted because complete coverage and native compilation are cold-start costs. Cache identity includes the program digest, AOT ABI, trace style, architecture, and operating system. Metadata and artifact digests reject stale or corrupt entries, and temporary-file replacement prevents partial artifacts from becoming visible.

Coverage is not treated as a correctness assumption. Unseen indirect targets, memory guards, ecalls, and exceptional paths continue in the interpreter. Fallback reasons remain active until execution reaches a compiled block leader, preserving recovery semantics and producing actionable diagnostics.

Change Highlights

• Replace bounded hot-root sampling with complete, low-overhead control-flow coverage.
• Load or build persistent preflight AOT artifacts through a versioned cache.
• Validate cache identity and artifact integrity before loading native code.
• Classify dynamic-PC, memory-guard, ecall, and exceptional fallbacks.
• Add tests for late targets, canonical block coverage, cache hits and corruption recovery, unseen indirect targets, and memory-guard recovery.

Benchmark / Performance Impact

Operation

Operation	baseline (s)	this PR (s)	Improvement
End-to-end	67.100	65.800	1.02x faster

Layer

Layer	baseline (s)	this PR (s)	Improvement
Preflight / emulator	6.500	5.510	1.18x faster (15.23% lower)
Application proving	50.700	50.100	1.01x faster
Recursion	9.160	9.420	0.97x
Root verification	0.028	0.028	1.00x

Benchmark command(s):

benchmark-dispatch: mainnet block 23817600

Environment: benchmark-dispatch runner and identical block input; branch-specific artifacts are linked below.

Raw data:

• baseline: resolve_oom/mainnet23817600
• this PR: full_aot_block/mainnet23817600

Testing

• Default and Goldilocks test suites pass in CI.
• Repository lint checks pass in CI.
• AOT tests cover full-workload roots, cache reuse, corrupt artifact and metadata rebuilds, runtime ABI offsets, dynamic-target fallback, and memory-guard recovery.

Risks and Rollout

• No circuit or protocol constraints change. A faulty AOT path can fail preflight or proving, but cannot relax proof constraints.
• Cached native artifacts increase local disk use; ABI-, program-, trace-, and platform-scoped keys prevent incompatible reuse.
• Interpreter fallback remains available for workload-dependent control flow. Per-reason fallback counters provide rollout diagnostics.
• Rollback consists of reverting the AOT cache path and clearing generated cache artifacts if required.

Follow-ups (optional)

• Add cache retention limits if artifact growth becomes operationally significant.

Copilot Reviewer Directive (keep this section)

When Copilot reviews this PR, apply .github/copilot-instructions.md strictly.