VM: remove legacy numeric recurrence body recognizers

Parent: #498 Related: #491, #465

The live quickening inventory labels pure and observable numeric recurrence paths as legacy_narrow / legacy_pattern_kernel. The pure path recognizes one exact bytecode body and computes a 256-element recurrence result/step table instead of executing the recursive calls; the observable variant recognizes one exact counter-updating body. Both conflict with the roadmap rule excluding body recognizers and computed-result caches.

Executed baseline: a no-native-owner fib(8) witness with VM inline calls disabled returns 21 and records 774 logical steps. Existing recurrence tests pass 4/4 and explicitly assert legacy-kernel hits. The VM-entry metric counts outer driver entry, so its value of 1 is not evidence of activation count; actual recursive statement events will be the call-observability witness.

Acceptance:

• Remove both recognizers/executors, their result/step tables, cached recurrence Chunk metadata, and test-only hit counters.
• Execute recursion through the general VM activation/call paths, preserving result 21, logical steps 774, rebinding, argument semantics, observable calls, and exceptions.
• Preserve and strengthen restriction/checkpoint, debugger/profiler, GC, no-GIL, termination, budget, and OOM witnesses using ordinary execution.
• Relevant exact before/after corpus and tree/VM comparisons introduce no new failures; focused normal, ReleaseSafe, and unsuppressed TSan gates pass.
• Remove live inventory coverage for the deleted kernels and regenerate docs/manifest; keep historical benchmark artifacts intact and identify the remaining legacy families honestly.

This removes benchmark-shaped work substitution. General no-JIT performance improvements remain open under #498; do not claim the removal establishes throughput parity.

Attribution correction: vm_entries counts outer driver entry, not each trampoline activation, so the observed value 1 must not be compared to 68 activations. The issue body is corrected. The executed legacy-kernel hit tests establish use of the recognizers; recursive statement-profiler events will verify ordinary call observability after removal. The result 21 and logical-step golden 774 remain valid measurements.

Landed in 7cbe01e. Both exact recurrence recognizers, the computed result/step table, cached Chunk plan, special executors, and hit counters are deleted. Recursion uses general VM calls and activations. Historical reports remain intact; the live inventory removes those two families and still identifies seven remaining legacy families.

Validation:

• recursive execution focused gate: 8/8 normal, unsuppressed TSan, and ReleaseSafe;
• result 21 and logical steps 774 retained; ordinary activation storage is exercised; profiler observes all 67 returns;
• expanded profiler/debugger/GC gate: 2/2 in all three modes; tree debugger observes all 67 returns and the existing required-bytecode debugger rejection remains explicit;
• restrictions after checkpoints, rebinding, observable counters/accessors, no-GIL workers, budget, termination, and GC reentry pass;
• activation OOM sweep extended with recursive execution: 2/2;
• exact before/after corpus remains 1,316 pass / 0 fail; paired output remains 1,313 agreements plus the same three pre-existing tree-fail/VM-pass TCO divergences, with byte-identical output for every group;
• inventory writer/self-test/validator, 378-file documentation links, and manifest-checked 51-page docs build pass.

Flips 0 test262 cases; no new measured failures or divergences. No throughput claim. Full hosted CI is pending. General debugger bytecode admission and the other legacy kernels remain broader roadmap work.