perf: hoist cap==0 check off full emplace

[equationzhao]

Summary

• Full emplace_back / emplace_front no longer call EQZ_COLD ensure_writable() on every overwrite. The helper runs only when capacity()==0 (still throws std::length_error).

• Closes the GCC 13 hot-path hole from Investigate GCC 13 bench gap versus Clang #1 without inlining throw into emplace_*.

• DESIGN.md matches the new shape and records the discarded R2/R3 variants (drop EQZ_COLD / always-inline check + cold throw helper), which sped up overwrite but sent GCC copy/full to ~11 µs.

• Adds docs/BENCH.md: keep-or-revert, local vs remote release-bench, and how to read GCC/Clang ratios. Issues and PR bodies are English (CONTRIBUTING.md, CODING_STANDARDS.md).

• Closes Investigate GCC 13 bench gap versus Clang #1

Test plan

• Existing unit tests: capacity()==0 push_back still throws length_error
• Experiment branch bench/r1-hoist-cap0 full CI matrix green
• Five-platform release-bench vs main 56fb5af:
  • GCC 13 x64: slide/u64/4097 5.02 ns → 0.61 ns; geomean 1.29 → 1.10 (0.90 if try_push/full timer-floor 469× is excluded)
  • macOS Clang geomean 0.567 → 0.569
  • Do not land R2/R3: same hot-path win, GCC copy/full 814 ns → ~11 µs
  • try_push/full on GCC going 0.01 ns → 0.70 ns (Boost still 0.00) poisons geomean; it is not an overwrite regression

[qodo-code-review]

ⓘ Qodo reviews are paused because the subscription is no longer active. Ask your workspace admin to reactivate the subscription to resume reviews. Manage billing