VM: build a fast no-JIT execution tier for restricted hosts

[chrisbbreuer]

Parent: #459
Depends on: #465, #461
Related: #466, #481, #500

Goal

Build a genuinely fast no-JIT execution tier for platforms, deployments, and security policies that cannot publish executable memory.

Current boundary

The tree-walker is the semantic baseline and the VM primarily provides suspend/resume, deep recursion, proper tail calls, and native-tier substrate. Existing quick kernels accelerate selected shapes, but there is no declared broad no-JIT performance contract. #465 makes correct code bytecode-eligible; this issue makes that bytecode competitive without executable mappings.

Scope

• Establish interpreter/VM attribution for every Bench: define the representative dependency-free performance matrix #460 workload with native tiers forcibly disabled.
• Implement general adaptive quickening, typed/specialized bytecodes, superinstructions, inline caches, threaded or otherwise efficient dispatch supported by Zig, activation/frame reuse, and narrow runtime helpers.
• Specialize property/index/call/arithmetic/control/allocation paths from runtime observations with exact versioned guards and dequickening.
• Preserve debugger/statement hooks, step/watchdog accounting, exceptions/finally, async/generator suspension, proper tail calls, moving GC roots, invalidation, OOM, and no-GIL synchronization.
• Bound quickened metadata/code growth and reset it safely on invalidation, realm teardown, memory pressure, or changed assumptions.
• Publish no-JIT cold/warm, code/metadata memory, latency, and 1/8-context evidence independently from JIT results.

No-workaround rules

• No source/function/benchmark recognizer, giant semantic-shape kernel, computed-result cache, disabled checks/hooks/GC, unsafe direct object storage, or fallback to a native executable mapping.
• A quick path must be a general bytecode/runtime specialization with exact miss and dequickening behavior.

Acceptance

• A machine-readable bytecode/quickening coverage and rejection inventory names every specialization and remaining dispatch family.
• Forced tree-walker/VM and quickening-off/on differentials match result, exception, effects, steps, jobs, debugger behavior, OOM, termination, and teardown.
• Moving GC, no-GIL, TSan, test262, PR-249, generators/async, threadfuzz, and fault gates pass with JIT unavailable.
• Perf: add causal attribution and exact-parent regression infrastructure #461 profiles prove dispatch/runtime reductions across predeclared Bench: define the representative dependency-free performance matrix #460 rows with bounded metadata/RSS.
• No-JIT performance and memory are published for every declared restricted platform; fallback is not counted as a native-tier success.

Foundation

Existing VM quickening and activation-stack work live in src/vm.zig; current architectural boundary is documented in docs/architecture.md and docs/advanced/execution-tiers.md.