perf: avoid concat scratch string interning

[ianm199]

Summary

• add a safe two-operand string/number concat fast path that formats numeric operands into the final concat buffer instead of interning a temporary numeric Lua string first
• add PROFILE_REPEAT=N support to hotspot, opcode, and GC profiling runners for short workload telemetry
• include gc_pressure, mandelbrot_long, and table_hash_pressure in the dashboard workload registry
• update performance docs with the concat finding and rejected VM/OP_MUL spikes

Evidence

Correctness:

• make test passed, including 44/44 official Lua conformance tests

Benchmarks, Apple M3 Max, best-of-N vs Lua 5.4.7 reference:

• focused best-of-10: table_hash_pressure 1.17x after the concat fast path (harness/bench/results/20260602T183040Z-98bd6bd-compare.tsv)
• full best-of-5: table_hash_pressure 1.14x, overall 1.54x (harness/bench/results/20260602T183215Z-98bd6bd-compare.tsv)

Profile delta:

• before: table_hash_pressure_x100 had lua_vm::vm::concat at 1,297 samples / 19.4% (harness/bench/profiles/20260602T182714Z-98bd6bd-table_hash_pressure_x100/summary.txt)
• after: lua_vm::vm::concat dropped to 116 samples / 3.1% (harness/bench/profiles/20260602T183359Z-98bd6bd-table_hash_pressure_x100/summary.txt)

Negative results recorded in docs:

• no-hook ci_trap guard looked plausible but regressed binarytrees in isolation
• direct-arm OP_MUL shape worsened mandelbrot_long, so it was dropped

Validation

• make test
• bash -n harness/bench/profile-hotspots.sh
• bash -n harness/bench/opcode-profile.sh
• bash -n harness/bench/gc-profile.sh
• python3 -m py_compile harness/bench/history.py harness/bench/vm-execute-attribution.py
• PROFILE_REPEAT=2 bash harness/bench/gc-profile.sh gc_pressure
• PROFILE_REPEAT=2 bash harness/bench/opcode-profile.sh table_hash_pressure
• python3 harness/bench/history.py --out-dir /tmp/lua-rs-history-check